Prepare new PyPI release.

* make clipnorm work with embeddings layer

* test for embedding + clipnorm

* update embedding_with_clipnorm test nb_epoch to epochs

.github/stale.yml

@@ -0,0 +1,19 @@
+# Configuration for probot-stale - https://github.com/probot/stale
+
+# Number of days of inactivity before an Issue or Pull Request becomes stale
+daysUntilStale: 90
+# Number of days of inactivity before a stale Issue or Pull Request is closed
+daysUntilClose: 30
+# Issues or Pull Requests with these labels will never be considered stale
+exemptLabels:
+  - bug
+  - Announcement
+  - help wanted
+  - To investigate
+# Label to use when marking as stale
+staleLabel: stale
+# Comment to post when marking as stale. Set to `false` to disable
+markComment: >
+  This issue has been automatically marked as stale because it has not had
+  recent activity. It will be closed after 30 days if no further activity 
+  occurs, but feel free to re-open a closed issue if needed.

.travis.yml

@@ -7,6 +7,8 @@ matrix:
           env: KERAS_BACKEND=tensorflow TEST_MODE=PEP8
         - python: 2.7
           env: KERAS_BACKEND=tensorflow TEST_MODE=INTEGRATION_TESTS
+        - python: 3.5
+          env: KERAS_BACKEND=tensorflow TEST_MODE=DOC
         - python: 2.7
           env: KERAS_BACKEND=tensorflow
         - python: 3.5
@@ -15,6 +17,10 @@ matrix:
           env: KERAS_BACKEND=theano
         - python: 3.5
           env: KERAS_BACKEND=theano
+        - python: 2.7
+          env: KERAS_BACKEND=cntk
+        - python: 3.5
+          env: KERAS_BACKEND=cntk
 install:
   # code below is taken from http://conda.pydata.org/docs/travis.html
   # We do this conditionally because it saves us some downloading if the
@@ -34,7 +40,7 @@ install:
 
   - conda create -q -n test-environment python=$TRAVIS_PYTHON_VERSION numpy scipy matplotlib pandas pytest h5py
   - source activate test-environment
-  - pip install git+git://github.com/Theano/Theano.git
+  - pip install theano
 
   # install PIL for preprocessing tests
   - if [[ "$TRAVIS_PYTHON_VERSION" == "2.7" ]]; then
@@ -45,8 +51,24 @@ install:
 
   - pip install -e .[tests]
 
-  # install TensorFlow
+  # install TensorFlow (CPU version).
   - pip install tensorflow
+  
+  # install cntk
+  - if [[ "$TRAVIS_PYTHON_VERSION" == "2.7" ]]; then
+      pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.0-cp27-cp27mu-linux_x86_64.whl;
+    elif [[ "$TRAVIS_PYTHON_VERSION" == "3.5" ]]; then
+      pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.0-cp35-cp35m-linux_x86_64.whl;
+    fi
+
+  #install open mpi
+  - rm -rf ~/mpi
+  - mkdir ~/mpi
+  - pushd ~/mpi
+  - wget http://cntk.ai/PythonWheel/ForKeras/depends/openmpi_1.10-3.zip
+  - unzip ./openmpi_1.10-3.zip
+  - sudo dpkg -i openmpi_1.10-3.deb
+  - popd
 
 # command to run tests
 script:
@@ -61,6 +83,8 @@ script:
        PYTHONPATH=$PWD:$PYTHONPATH py.test tests/integration_tests;
     elif [[ "$TEST_MODE" == "PEP8" ]]; then
        PYTHONPATH=$PWD:$PYTHONPATH py.test --pep8 -m pep8 -n0;
+    elif [[ "$TEST_MODE" == "DOC" ]]; then
+        PYTHONPATH=$PWD:$PYTHONPATH py.test tests/test_documentation.py;
     else
-       PYTHONPATH=$PWD:$PYTHONPATH py.test tests/ --ignore=tests/integration_tests --cov=keras tests/ --cov-fail-under 78 --cov-report term-missing;
+       PYTHONPATH=$PWD:$PYTHONPATH py.test tests/ --ignore=tests/integration_tests --ignore=tests/test_documentation.py --cov=keras tests/ --cov-fail-under 78 --cov-report term-missing;
     fi

CONTRIBUTING.md

@@ -19,6 +19,7 @@ To easily update Theano: `pip install git+git://github.com/Theano/Theano.git --u
 
 The more information you provide, the easier it is for us to validate that there is a bug and the faster we'll be able to take action. If you want your issue to be resolved quickly, following the steps above is crucial.
 
+---
 
 ## Requesting a Feature
 
@@ -31,11 +32,15 @@ You can also use Github issues to request features you would like to see in Kera
 3. After discussing the feature you may choose to attempt a Pull Request. If you're at all able, start writing some code. We always have more work to do than time to do it. If you can write some code then that will speed the process along.
 
 
+---
+
 ## Requests for Contributions
 
 [This is the board](https://github.com/fchollet/keras/projects/1) where we list current outstanding issues and features to be added. If you want to start contributing to Keras, this is the place to start.
 
 
+---
+
 ## Pull Requests
 
 **Where should I submit my pull request?**
@@ -49,27 +54,29 @@ Here's a quick guide to submitting your improvements:
 
 2. Write the code. This is the hard part!
 
-3. Make sure any new function or class you introduce has proper docstrings. Make sure any code you touch still has up-to-date docstrings and documentation.
+3. Make sure any new function or class you introduce has proper docstrings. Make sure any code you touch still has up-to-date docstrings and documentation. **Docstring style should be respected.** In particular, they should be formatted in MarkDown, and there should be sections for `Arguments`, `Returns`, `Raises` (if applicable). Look at other docstrings in the codebase for examples.
 
 4. Write tests. Your code should have full unit test coverage. If you want to see your PR merged promptly, this is crucial.
 
 5. Run our test suite locally. It's easy: from the Keras folder, simply run: `py.test tests/`.
-  - You will need to install the test requirements as well: `pip install -e .[tests]`.
+    - You will need to install the test requirements as well: `pip install -e .[tests]`.
 
 6. Make sure all tests are passing:
-  - with the Theano backend, on Python 2.7 and Python 3.5
-  - with the TensorFlow backend, on Python 2.7
+    - with the Theano backend, on Python 2.7 and Python 3.5. Make sure you have the development version of Theano.
+    - with the TensorFlow backend, on Python 2.7 and Python 3.5. Make sure you have the development version of TensorFlow.
 
 7. We use PEP8 syntax conventions, but we aren't dogmatic when it comes to line length. Make sure your lines stay reasonably sized, though. To make your life easier, we recommend running a PEP8 linter:
-  - Install PEP8 packages: `pip install pep8 pytest-pep8 autopep8`
-  - Run a standalone PEP8 check: `py.test --pep8 -m pep8`
-  - You can automatically fix some PEP8 error by running: `autopep8 -i --select <errors> <FILENAME>` for example: `autopep8 -i --select E128 tests/keras/backend/test_backends.py`
+    - Install PEP8 packages: `pip install pep8 pytest-pep8 autopep8`
+    - Run a standalone PEP8 check: `py.test --pep8 -m pep8`
+    - You can automatically fix some PEP8 error by running: `autopep8 -i --select <errors> <FILENAME>` for example: `autopep8 -i --select E128 tests/keras/backend/test_backends.py`
 
-8. When committing, use appropriate, descriptive commit messages. Make sure that your branch history is not a string of "bug fix", "fix", "oops", etc. When submitting your PR, squash your commits into a single commit with an appropriate commit message, to make sure the project history stays clean and readable. See ['rebase and squash'](http://rebaseandsqua.sh/) for technical help on how to squash your commits.
+8. When committing, use appropriate, descriptive commit messages.
 
 9. Update the documentation. If introducing new functionality, make sure you include code snippets demonstrating the usage of your new feature.
 
-10. Submit your PR. If your changes have been approved in a previous discussion, and if you have complete (and passing) unit tests, your PR is likely to be merged promptly. Otherwise, well...
+10. Submit your PR. If your changes have been approved in a previous discussion, and if you have complete (and passing) unit tests as well as proper docstrings/documentation, your PR is likely to be merged promptly. Otherwise, well...
+
+---
 
 ## Adding new examples
 

LICENSE

@@ -8,8 +8,12 @@ All contributions by Google:
 Copyright (c) 2015, Google, Inc.
 All rights reserved.
 
+All contributions by Microsoft:
+Copyright (c) 2017, Microsoft, Inc.
+All rights reserved.
+
 All other contributions:
-Copyright (c) 2015, the respective contributors.
+Copyright (c) 2015 - 2017, the respective contributors.
 All rights reserved.
 
 Each contributor holds copyright over their respective contributions.

README.md

@@ -1,11 +1,11 @@
-# Keras: Deep Learning library for TensorFlow and Theano
+# Keras: Deep Learning for Python
 
 [![Build Status](https://travis-ci.org/fchollet/keras.svg?branch=master)](https://travis-ci.org/fchollet/keras)
 [![license](https://img.shields.io/github/license/mashape/apistatus.svg?maxAge=2592000)](https://github.com/fchollet/keras/blob/master/LICENSE)
 
 ## You have just found Keras.
 
-Keras is a high-level neural networks API, written in Python and capable of running on top of either [TensorFlow](https://github.com/tensorflow/tensorflow) or [Theano](https://github.com/Theano/Theano). It was developed with a focus on enabling fast experimentation. *Being able to go from idea to result with the least possible delay is key to doing good research.*
+Keras is a high-level neural networks API, written in Python and capable of running on top of either [TensorFlow](https://github.com/tensorflow/tensorflow), [CNTK](https://github.com/Microsoft/cntk) or [Theano](https://github.com/Theano/Theano). It was developed with a focus on enabling fast experimentation. *Being able to go from idea to result with the least possible delay is key to doing good research.*
 
 Use Keras if you need a deep learning library that:
 
@@ -125,6 +125,11 @@ Keras uses the following dependencies:
 - TensorFlow
     - [See installation instructions](https://www.tensorflow.org/install/).
 
+*When using the CNTK backend:*
+
+- CNTK
+    - [See installation instructions](https://docs.microsoft.com/en-us/cognitive-toolkit/setup-cntk-on-your-machine).
+
 *When using the Theano backend:*
 
 - Theano
@@ -143,7 +148,7 @@ sudo pip install keras
 ------------------
 
 
-## Switching from TensorFlow to Theano
+## Switching from TensorFlow to CNTK or Theano
 
 By default, Keras will use TensorFlow as its tensor manipulation library. [Follow these instructions](http://keras.io/backend/) to configure the Keras backend.
 

docs/autogen.py

@@ -8,9 +8,7 @@ Index
 - Getting started
     Getting started with the sequential model
     Getting started with the functional api
-    Examples
     FAQ
-    Installation guide
 
 - Models
     About Keras models
@@ -26,18 +24,23 @@ Index
         explain common layer functions: get_weights, set_weights, get_config
         explain input_shape
         explain usage on non-Keras tensors
-    Core layers
-    Convolutional
-    Recurrent
-    Embeddings
-    Normalization
-    Advanced activations
-    Noise
+    Core Layers
+    Convolutional Layers
+    Pooling Layers
+    Locally-connected Layers
+    Recurrent Layers
+    Embedding Layers
+    Merge Layers
+    Advanced Activations Layers
+    Normalization Layers
+    Noise Layers
+    Layer Wrappers
+    Writing your own Keras layers
 
 - Preprocessing
-    Image preprocessing
-    Text preprocessing
-    Sequence preprocessing
+    Sequence Preprocessing
+    Text Preprocessing
+    Image Preprocessing
 
 Losses
 Metrics
@@ -45,12 +48,15 @@ Optimizers
 Activations
 Callbacks
 Datasets
+Applications
 Backend
-Initializations
+Initializers
 Regularizers
 Constraints
 Visualization
 Scikit-learn API
+Utils
+Contributing
 
 '''
 from __future__ import print_function
@@ -114,14 +120,13 @@ PAGES = [
             models.Sequential.fit,
             models.Sequential.evaluate,
             models.Sequential.predict,
-            models.Sequential.predict_classes,
-            models.Sequential.predict_proba,
             models.Sequential.train_on_batch,
             models.Sequential.test_on_batch,
             models.Sequential.predict_on_batch,
             models.Sequential.fit_generator,
             models.Sequential.evaluate_generator,
             models.Sequential.predict_generator,
+            models.Sequential.get_layer,
         ],
     },
     {
@@ -382,7 +387,7 @@ def process_class_docstring(docstring):
                        r'\n    __\1__\n\n',
                        docstring)
 
-    docstring = re.sub(r'    ([^\s\\]+):(.*)\n',
+    docstring = re.sub(r'    ([^\s\\\(]+):(.*)\n',
                        r'    - __\1__:\2\n',
                        docstring)
 
@@ -400,7 +405,7 @@ def process_function_docstring(docstring):
                        r'\n        __\1__\n\n',
                        docstring)
 
-    docstring = re.sub(r'    ([^\s\\]+):(.*)\n',
+    docstring = re.sub(r'    ([^\s\\\(]+):(.*)\n',
                        r'    - __\1__:\2\n',
                        docstring)
 
@@ -509,3 +514,5 @@ for page_data in PAGES:
     if not os.path.exists(subdir):
         os.makedirs(subdir)
     open(path, 'w').write(mkdown)
+
+shutil.copyfile('../CONTRIBUTING.md', 'sources/contributing.md')

docs/mkdocs.yml

@@ -51,3 +51,4 @@ pages:
 - Visualization: visualization.md
 - Scikit-learn API: scikit-learn-api.md
 - Utils: utils.md
+- Contributing: contributing.md

docs/templates/applications.md

@@ -15,7 +15,7 @@ Weights are downloaded automatically when instantiating a model. They are stored
 - [ResNet50](#resnet50)
 - [InceptionV3](#inceptionv3)
 
-All of these architectures (except Xception) are compatible with both TensorFlow and Theano, and upon instantiation the models will be built according to the image data format set in your Keras configuration file at `~/.keras/keras.json`. For instance, if you have set `image_data_format=tf`, then any model loaded from this repository will get built according to the TensorFlow data format convention, "Width-Height-Depth".
+All of these architectures (except Xception) are compatible with both TensorFlow and Theano, and upon instantiation the models will be built according to the image data format set in your Keras configuration file at `~/.keras/keras.json`. For instance, if you have set `image_data_format=channels_last`, then any model loaded from this repository will get built according to the TensorFlow data format convention, "Width-Height-Depth".
 
 The Xception model is only available for TensorFlow, due to its reliance on `SeparableConvolution` layers.
 
@@ -130,10 +130,10 @@ for i, layer in enumerate(base_model.layers):
    print(i, layer.name)
 
 # we chose to train the top 2 inception blocks, i.e. we will freeze
-# the first 172 layers and unfreeze the rest:
-for layer in model.layers[:172]:
+# the first 249 layers and unfreeze the rest:
+for layer in model.layers[:249]:
    layer.trainable = False
-for layer in model.layers[172:]:
+for layer in model.layers[249:]:
    layer.trainable = True
 
 # we need to recompile the model for these modifications to take effect
@@ -253,7 +253,7 @@ The default input size for this model is 224x224.
 - input_shape: optional shape tuple, only to be specified
     if `include_top` is False (otherwise the input shape
     has to be `(224, 224, 3)` (with `channels_last` data format)
-    or `(3, 224, 244)` (with `channels_first` data format).
+    or `(3, 224, 224)` (with `channels_first` data format).
     It should have exactly 3 inputs channels,
     and width and height should be no smaller than 48.
     E.g. `(200, 200, 3)` would be one valid value.
@@ -309,7 +309,7 @@ The default input size for this model is 224x224.
 - input_shape: optional shape tuple, only to be specified
     if `include_top` is False (otherwise the input shape
     has to be `(224, 224, 3)` (with `channels_last` data format)
-    or `(3, 224, 244)` (with `channels_first` data format).
+    or `(3, 224, 224)` (with `channels_first` data format).
     It should have exactly 3 inputs channels,
     and width and height should be no smaller than 48.
     E.g. `(200, 200, 3)` would be one valid value.
@@ -367,7 +367,7 @@ The default input size for this model is 224x224.
 - input_shape: optional shape tuple, only to be specified
     if `include_top` is False (otherwise the input shape
     has to be `(224, 224, 3)` (with `channels_last` data format)
-    or `(3, 224, 244)` (with `channels_first` data format).
+    or `(3, 224, 224)` (with `channels_first` data format).
     It should have exactly 3 inputs channels,
     and width and height should be no smaller than 197.
     E.g. `(200, 200, 3)` would be one valid value.

docs/templates/backend.md

@@ -4,12 +4,13 @@
 
 Keras is a model-level library, providing high-level building blocks for developing deep learning models. It does not handle itself low-level operations such as tensor products, convolutions and so on. Instead, it relies on a specialized, well-optimized tensor manipulation library to do so, serving as the "backend engine" of Keras. Rather than picking one single tensor library and making the implementation of Keras tied to that library, Keras handles the problem in a modular way, and several different backend engines can be plugged seamlessly into Keras.
 
-At this time, Keras has two backend implementations available: the **TensorFlow** backend and the **Theano** backend.
+At this time, Keras has three backend implementations available: the **TensorFlow** backend, the **Theano** backend, and the **CNTK** backend.
 
 - [TensorFlow](http://www.tensorflow.org/) is an open-source symbolic tensor manipulation framework developed by Google, Inc.
 - [Theano](http://deeplearning.net/software/theano/) is an open-source symbolic tensor manipulation framework developed by LISA/MILA Lab at Université de Montréal.
+- [CNTK](https://www.microsoft.com/en-us/cognitive-toolkit/) is an open-source, commercial-grade toolkit for deep learning developed by Microsoft.
 
-In the future, we are likely to add more backend options. Go ask Microsoft about how their CNTK backend project is doing.
+In the future, we are likely to add more backend options.
 
 ----
 
@@ -34,7 +35,7 @@ The default configuration file looks like this:
 }
 ```
 
-Simply change the field `backend` to either `"theano"` or `"tensorflow"`, and Keras will use the new configuration next time you run any Keras code.
+Simply change the field `backend` to `"theano"`, `"tensorflow"`, or `"cntk"`, and Keras will use the new configuration next time you run any Keras code.
 
 You can also define the environment variable ``KERAS_BACKEND`` and this will
 override what is defined in your config file :
@@ -65,7 +66,7 @@ You can change these settings by editing `$HOME/.keras/keras.json`.
   - For 3D data, `"channels_last"` assumes `(conv_dim1, conv_dim2, conv_dim3, channels)` while `"channels_first"` assumes `(channels, conv_dim1, conv_dim2, conv_dim3)`.
 * `epsilon`: float, a numeric fuzzing constant used to avoid dividing by zero in some operations.
 * `floatx`: string, `"float16"`, `"float32"`, or `"float64"`. Default float precision.
-* `backend`: string, `"tensorflow"` or `"theano"`.
+* `backend`: string, `"tensorflow"`, `"theano"`, or `"cntk"`.
 
 ----
 

docs/templates/callbacks.md

@@ -36,14 +36,14 @@ class LossHistory(keras.callbacks.Callback):
         self.losses.append(logs.get('loss'))
 
 model = Sequential()
-model.add(Dense(10, input_dim=784, init='uniform'))
+model.add(Dense(10, input_dim=784, kernel_initializer='uniform'))
 model.add(Activation('softmax'))
 model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
 
 history = LossHistory()
-model.fit(X_train, Y_train, batch_size=128, epochs=20, verbose=0, callbacks=[history])
+model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=0, callbacks=[history])
 
-print history.losses
+print(history.losses)
 # outputs
 '''
 [0.66047596406559383, 0.3547245744908703, ..., 0.25953155204159617, 0.25901699725311789]
@@ -58,15 +58,13 @@ print history.losses
 from keras.callbacks import ModelCheckpoint
 
 model = Sequential()
-model.add(Dense(10, input_dim=784, init='uniform'))
+model.add(Dense(10, input_dim=784, kernel_initializer='uniform'))
 model.add(Activation('softmax'))
 model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
 
 '''
 saves the model weights after each epoch if the validation loss decreased
 '''
-checkpointer = ModelCheckpoint(filepath="/tmp/weights.hdf5", verbose=1, save_best_only=True)
-model.fit(X_train, Y_train, batch_size=128, epochs=20, verbose=0, validation_data=(X_test, Y_test), callbacks=[checkpointer])
-
+checkpointer = ModelCheckpoint(filepath='/tmp/weights.hdf5', verbose=1, save_best_only=True)
+model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=0, validation_data=(X_test, Y_test), callbacks=[checkpointer])
 ```
-

docs/templates/datasets.md

@@ -55,7 +55,7 @@ As a convention, "0" does not stand for a specific word, but instead is used to
 ```python
 from keras.datasets import imdb
 
-(x_train, y_train), (x_test, y_test) = imdb.load_data(path="imdb_full.pkl",
+(x_train, y_train), (x_test, y_test) = imdb.load_data(path="imdb.npz",
                                                       num_words=None,
                                                       skip_top=0,
                                                       maxlen=None,
@@ -72,13 +72,13 @@ from keras.datasets import imdb
 - __Arguments:__
 
     - __path__: if you do not have the data locally (at `'~/.keras/datasets/' + path`), it will be downloaded to this location.
-    - __num_words__: integer or None. Top most frequent words to consider. Any less frequent word will appear as 0 in the sequence data.
-    - __skip_top__: integer. Top most frequent words to ignore (they will appear as 0s in the sequence data).
+    - __num_words__: integer or None. Top most frequent words to consider. Any less frequent word will appear as `oov_char` value in the sequence data.
+    - __skip_top__: integer. Top most frequent words to ignore (they will appear as `oov_char` value in the sequence data).
     - __maxlen__: int. Maximum sequence length. Any longer sequence will be truncated.
     - __seed__: int. Seed for reproducible data shuffling.
-    - __start_char__: char. The start of a sequence will be marked with this character.
+    - __start_char__: int. The start of a sequence will be marked with this character.
         Set to 1 because 0 is usually the padding character.
-    - __oov_char__: char. words that were cut out because of the `num_words`
+    - __oov_char__: int. words that were cut out because of the `num_words`
         or `skip_top` limit will be replaced with this character.
     - __index_from__: int. Index actual words with this index and higher.
 
@@ -94,7 +94,7 @@ Dataset of 11,228 newswires from Reuters, labeled over 46 topics. As with the IM
 ```python
 from keras.datasets import reuters
 
-(x_train, y_train), (x_test, y_test) = reuters.load_data(path="reuters.pkl",
+(x_train, y_train), (x_test, y_test) = reuters.load_data(path="reuters.npz",
                                                          num_words=None,
                                                          skip_top=0,
                                                          maxlen=None,
@@ -107,12 +107,12 @@ from keras.datasets import reuters
 
 The specifications are the same as that of the IMDB dataset, with the addition of:
 
-    - __test_split__: float. Fraction of the dataset to be used as test data.
+- __test_split__: float. Fraction of the dataset to be used as test data.
 
 This dataset also makes available the word index used for encoding the sequences:
 
 ```python
-word_index = reuters.get_word_index(path="reuters_word_index.pkl")
+word_index = reuters.get_word_index(path="reuters_word_index.json")
 ```
 
 - __Returns:__ A dictionary where key are words (str) and values are indexes (integer). eg. `word_index["giraffe"]` might return `1234`. 

docs/templates/getting-started/faq.md

@@ -27,7 +27,7 @@ Please cite Keras in your publications if it helps your research. Here is an exa
 ```
 @misc{chollet2015keras,
   title={Keras},
-  author={Chollet, Fran\c{c}ois},
+  author={Chollet, Fran\c{c}ois and others},
   year={2015},
   publisher={GitHub},
   howpublished={\url{https://github.com/fchollet/keras}},
@@ -38,7 +38,7 @@ Please cite Keras in your publications if it helps your research. Here is an exa
 
 ### How can I run Keras on GPU?
 
-If you are running on the TensorFlow backend, your code will automatically run on GPU if any available GPU is detected.
+If you are running on the TensorFlow or CNTK backends, your code will automatically run on GPU if any available GPU is detected.
 
 If you are running on the Theano backend, you can use one of the following methods:
 
@@ -153,16 +153,16 @@ For example:
 """
 Assume original model looks like this:
     model = Sequential()
-    model.add(Dense(2, input_dim=3, name="dense_1"))
-    model.add(Dense(3, name="dense_2"))
+    model.add(Dense(2, input_dim=3, name='dense_1'))
+    model.add(Dense(3, name='dense_2'))
     ...
     model.save_weights(fname)
 """
 
 # new model
 model = Sequential()
-model.add(Dense(2, input_dim=3, name="dense_1"))  # will be loaded
-model.add(Dense(10, name="new_dense"))  # will not be loaded
+model.add(Dense(2, input_dim=3, name='dense_1'))  # will be loaded
+model.add(Dense(10, name='new_dense'))  # will not be loaded
 
 # load weights from first model; will only affect the first layer, dense_1.
 model.load_weights(fname, by_name=True)
@@ -201,7 +201,7 @@ from keras import backend as K
 # with a Sequential model
 get_3rd_layer_output = K.function([model.layers[0].input],
                                   [model.layers[3].output])
-layer_output = get_3rd_layer_output([X])[0]
+layer_output = get_3rd_layer_output([x])[0]
 ```
 
 Similarly, you could build a Theano and TensorFlow function directly.
@@ -214,17 +214,17 @@ get_3rd_layer_output = K.function([model.layers[0].input, K.learning_phase()],
                                   [model.layers[3].output])
 
 # output in test mode = 0
-layer_output = get_3rd_layer_output([X, 0])[0]
+layer_output = get_3rd_layer_output([x, 0])[0]
 
 # output in train mode = 1
-layer_output = get_3rd_layer_output([X, 1])[0]
+layer_output = get_3rd_layer_output([x, 1])[0]
 ```
 
 ---
 
 ### How can I use Keras with datasets that don't fit in memory?
 
-You can do batch training using `model.train_on_batch(X, y)` and `model.test_on_batch(X, y)`. See the [models documentation](/models/sequential).
+You can do batch training using `model.train_on_batch(x, y)` and `model.test_on_batch(x, y)`. See the [models documentation](/models/sequential).
 
 Alternatively, you can write a generator that yields batches of training data and use the method `model.fit_generator(data_generator, steps_per_epoch, epochs)`.
 
@@ -239,7 +239,7 @@ You can use an `EarlyStopping` callback:
 ```python
 from keras.callbacks import EarlyStopping
 early_stopping = EarlyStopping(monitor='val_loss', patience=2)
-model.fit(X, y, validation_split=0.2, callbacks=[early_stopping])
+model.fit(x, y, validation_split=0.2, callbacks=[early_stopping])
 ```
 
 Find out more in the [callbacks documentation](/callbacks).
@@ -268,7 +268,7 @@ Validation data is never shuffled.
 The `model.fit` method returns an `History` callback, which has a `history` attribute containing the lists of successive losses and other metrics.
 
 ```python
-hist = model.fit(X, y, validation_split=0.2)
+hist = model.fit(x, y, validation_split=0.2)
 print(hist.history)
 ```
 
@@ -315,7 +315,7 @@ Making a RNN stateful means that the states for the samples of each batch will b
 When using stateful RNNs, it is therefore assumed that:
 
 - all batches have the same number of samples
-- If `X1` and `X2` are successive batches of samples, then `X2[i]` is the follow-up sequence to `X1[i]`, for every `i`.
+- If `x1` and `x2` are successive batches of samples, then `x2[i]` is the follow-up sequence to `x1[i]`, for every `i`.
 
 To use statefulness in RNNs, you need to:
 
@@ -332,7 +332,7 @@ Example:
 
 ```python
 
-X  # this is our input data, of shape (32, 21, 16)
+x  # this is our input data, of shape (32, 21, 16)
 # we will feed it to our model in sequences of length 10
 
 model = Sequential()
@@ -342,10 +342,10 @@ model.add(Dense(16, activation='softmax'))
 model.compile(optimizer='rmsprop', loss='categorical_crossentropy')
 
 # we train the network to predict the 11th timestep given the first 10:
-model.train_on_batch(X[:, :10, :], np.reshape(X[:, 10, :], (32, 16)))
+model.train_on_batch(x[:, :10, :], np.reshape(x[:, 10, :], (32, 16)))
 
 # the state of the network has changed. We can feed the follow-up sequences:
-model.train_on_batch(X[:, 10:20, :], np.reshape(X[:, 20, :], (32, 16)))
+model.train_on_batch(x[:, 10:20, :], np.reshape(x[:, 20, :], (32, 16)))
 
 # let's reset the states of the LSTM layer:
 model.reset_states()
@@ -411,15 +411,15 @@ The VGG16 model is also the basis for several Keras example scripts:
 
 ### How can I use HDF5 inputs with Keras?
 
-You can use the `HDF5Matrix` class from `keras.utils.io_utils`. See [the HDF5Matrix documentation](/io_utils/#HDF5Matrix) for details.
+You can use the `HDF5Matrix` class from `keras.utils.io_utils`. See [the HDF5Matrix documentation](/utils/#hdf5matrix) for details.
 
 You can also directly use a HDF5 dataset:
 
 ```python
 import h5py
 with h5py.File('input/file.hdf5', 'r') as f:
-    X_data = f['X_data']
-    model.predict(X_data)
+    x_data = f['x_data']
+    model.predict(x_data)
 ```
 
 ---
@@ -451,6 +451,6 @@ It contains the following fields:
 - The image data format to be used as default by image processing layers and utilities (either `channels_last` or `channels_first`).
 - The `epsilon` numerical fuzz factor to be used to prevent division by zero in some operations.
 - The default float data type.
-- The default backend. See the (backend documentation)[/backend].
+- The default backend. See the [backend documentation](/backend).
 
 Likewise, cached dataset files, such as those downloaded with [`get_file()`](/utils/#get_file), are stored by default in `$HOME/.keras/datasets/`.

docs/templates/getting-started/functional-api-guide.md

@@ -361,7 +361,7 @@ from keras.models import Model, Sequential
 # First, let's define a vision model using a Sequential model.
 # This model will encode an image into a vector.
 vision_model = Sequential()
-vision_model.add(Conv2D(64, (3, 3) activation='relu', padding='same', input_shape=(3, 224, 224)))
+vision_model.add(Conv2D(64, (3, 3), activation='relu', padding='same', input_shape=(3, 224, 224)))
 vision_model.add(Conv2D(64, (3, 3), activation='relu'))
 vision_model.add(MaxPooling2D((2, 2)))
 vision_model.add(Conv2D(128, (3, 3), activation='relu', padding='same'))

docs/templates/getting-started/sequential-model-guide.md

@@ -354,7 +354,7 @@ A stateful recurrent model is one for which the internal states (memories) obtai
 of samples are reused as initial states for the samples of the next batch. This allows to process longer sequences
 while keeping computational complexity manageable.
 
-[You can read more about stateful RNNs in the FAQ.](/faq/#how-can-i-use-stateful-rnns)
+[You can read more about stateful RNNs in the FAQ.](/getting-started/faq/#how-can-i-use-stateful-rnns)
 
 ```python
 from keras.models import Sequential

docs/templates/index.md

@@ -1,3 +1,3 @@
-# Keras: Deep Learning library for Theano and TensorFlow
+# Keras: The Python Deep Learning library
 
 {{autogenerated}}

docs/templates/initializers.md

@@ -39,5 +39,5 @@ from keras import backend as K
 def my_init(shape, dtype=None):
     return K.random_normal(shape, dtype=dtype)
 
-model.add(Dense(64, init=my_init))
+model.add(Dense(64, kernel_initializer=my_init))
 ```

docs/templates/layers/writing-your-own-keras-layers.md

@@ -21,7 +21,8 @@ class MyLayer(Layer):
 
     def build(self, input_shape):
         # Create a trainable weight variable for this layer.
-        self.kernel = self.add_weight(shape=(input_shape[1], self.output_dim),
+        self.kernel = self.add_weight(name='kernel', 
+                                      shape=(input_shape[1], self.output_dim),
                                       initializer='uniform',
                                       trainable=True)
         super(MyLayer, self).build(input_shape)  # Be sure to call this somewhere!

docs/templates/preprocessing/image.md

@@ -7,6 +7,7 @@ keras.preprocessing.image.ImageDataGenerator(featurewise_center=False,
     featurewise_std_normalization=False,
     samplewise_std_normalization=False,
     zca_whitening=False,
+    zca_epsilon=1e-6,
     rotation_range=0.,
     width_shift_range=0.,
     height_shift_range=0.,
@@ -29,6 +30,7 @@ Generate batches of tensor image data with real-time data augmentation. The data
     - __samplewise_center__: Boolean. Set each sample mean to 0.
     - __featurewise_std_normalization__: Boolean. Divide inputs by std of the dataset, feature-wise.
     - __samplewise_std_normalization__: Boolean. Divide each input by its std.
+    - __zca_epsilon__: epsilon for ZCA whitening. Default is 1e-6.
     - __zca_whitening__: Boolean. Apply ZCA whitening.
     - __rotation_range__: Int. Degree range for random rotations.
     - __width_shift_range__: Float (fraction of total width). Range for random horizontal shifts.
@@ -56,19 +58,19 @@ Generate batches of tensor image data with real-time data augmentation. The data
         If you never set it, then it will be "channels_last".
 
 - __Methods__:
-    - __fit(X)__: Compute the internal data stats related to the data-dependent transformations, based on an array of sample data.
+    - __fit(x)__: Compute the internal data stats related to the data-dependent transformations, based on an array of sample data.
         Only required if featurewise_center or featurewise_std_normalization or zca_whitening.
         - __Arguments__:
-            - __X__: sample data. Should have rank 4.
+            - __x__: sample data. Should have rank 4.
                 In case of grayscale data,
                 the channels axis should have value 1, and in case
                 of RGB data, it should have value 3.
             - __augment__: Boolean (default: False). Whether to fit on randomly augmented samples.
             - __rounds__: int (default: 1). If augment, how many augmentation passes over the data to use.
             - __seed__: int (default: None). Random seed.
-    - __flow(X, y)__: Takes numpy data & label arrays, and generates batches of augmented/normalized data. Yields batches indefinitely, in an infinite loop.
+    - __flow(x, y)__: Takes numpy data & label arrays, and generates batches of augmented/normalized data. Yields batches indefinitely, in an infinite loop.
         - __Arguments__:
-            - __X__: data. Should have rank 4.
+            - __x__: data. Should have rank 4.
                 In case of grayscale data,
                 the channels axis should have value 1, and in case
                 of RGB data, it should have value 3.
@@ -78,7 +80,7 @@ Generate batches of tensor image data with real-time data augmentation. The data
             - __seed__: int (default: None).
             - __save_to_dir__: None or str (default: None). This allows you to optimally specify a directory to which to save the augmented pictures being generated (useful for visualizing what you are doing).
             - __save_prefix__: str (default: `''`). Prefix to use for filenames of saved pictures (only relevant if `save_to_dir` is set).
-            - __save_format__: one of "png", "jpeg" (only relevant if `save_to_dir` is set). Default: "jpeg".
+            - __save_format__: one of "png", "jpeg" (only relevant if `save_to_dir` is set). Default: "png".
         - __yields__: Tuples of `(x, y)` where `x` is a numpy array of image data and `y` is a numpy array of corresponding labels.
             The generator loops indefinitely.
     - __flow_from_directory(directory)__: Takes the path to a directory, and generates batches of augmented/normalized data. Yields batches indefinitely, in an infinite loop.
@@ -88,25 +90,25 @@ Generate batches of tensor image data with real-time data augmentation. The data
                 See [this script](https://gist.github.com/fchollet/0830affa1f7f19fd47b06d4cf89ed44d) for more details.
             - __target_size__: tuple of integers, default: `(256, 256)`. The dimensions to which all images found will be resized.
             - __color_mode__: one of "grayscale", "rbg". Default: "rgb". Whether the images will be converted to have 1 or 3 color channels.
-            - __classes__: optional list of class subdirectories (e.g. `['dogs', 'cats']`). Default: None. If not provided, the list of classes will be automatically inferred (and the order of the classes, which will map to the label indices, will be alphanumeric).
-            - __class_mode__: one of "categorical", "binary", "sparse" or None. Default: "categorical". Determines the type of label arrays that are returned: "categorical" will be 2D one-hot encoded labels, "binary" will be 1D binary labels, "sparse" will be 1D integer labels. If None, no labels are returned (the generator will only yield batches of image data, which is useful to use `model.predict_generator()`, `model.evaluate_generator()`, etc.).
+            - __classes__: optional list of class subdirectories (e.g. `['dogs', 'cats']`). Default: None. If not provided, the list of classes will be automatically inferred from the subdirectory names/structure under `directory`, where each subdirectory will be treated as a different class (and the order of the classes, which will map to the label indices, will be alphanumeric). The dictionary containing the mapping from class names to class indices can be obtained via the attribute `class_indices`.
+            - __class_mode__: one of "categorical", "binary", "sparse" or None. Default: "categorical". Determines the type of label arrays that are returned: "categorical" will be 2D one-hot encoded labels, "binary" will be 1D binary labels, "sparse" will be 1D integer labels. If None, no labels are returned (the generator will only yield batches of image data, which is useful to use `model.predict_generator()`, `model.evaluate_generator()`, etc.). Please note that in case of class_mode None, the data still needs to reside in a subdirectory of `directory` for it to work correctly.
             - __batch_size__: size of the batches of data (default: 32).
             - __shuffle__: whether to shuffle the data (default: True)
             - __seed__: optional random seed for shuffling and transformations.
             - __save_to_dir__: None or str (default: None). This allows you to optimally specify a directory to which to save the augmented pictures being generated (useful for visualizing what you are doing).
             - __save_prefix__: str. Prefix to use for filenames of saved pictures (only relevant if `save_to_dir` is set).
-            - __save_format__: one of "png", "jpeg" (only relevant if `save_to_dir` is set). Default: "jpeg".
+            - __save_format__: one of "png", "jpeg" (only relevant if `save_to_dir` is set). Default: "png".
             - __follow_links__: whether to follow symlinks inside class subdirectories (default: False).
 
 
 - __Examples__:
 
-Example of using `.flow(X, y)`:
+Example of using `.flow(x, y)`:
 
 ```python
-(X_train, y_train), (X_test, y_test) = cifar10.load_data()
-Y_train = np_utils.to_categorical(y_train, num_classes)
-Y_test = np_utils.to_categorical(y_test, num_classes)
+(x_train, y_train), (x_test, y_test) = cifar10.load_data()
+y_train = np_utils.to_categorical(y_train, num_classes)
+y_test = np_utils.to_categorical(y_test, num_classes)
 
 datagen = ImageDataGenerator(
     featurewise_center=True,
@@ -118,20 +120,20 @@ datagen = ImageDataGenerator(
 
 # compute quantities required for featurewise normalization
 # (std, mean, and principal components if ZCA whitening is applied)
-datagen.fit(X_train)
+datagen.fit(x_train)
 
 # fits the model on batches with real-time data augmentation:
-model.fit_generator(datagen.flow(X_train, Y_train, batch_size=32),
-                    steps_per_epoch=len(X_train), epochs=epochs)
+model.fit_generator(datagen.flow(x_train, y_train, batch_size=32),
+                    steps_per_epoch=len(x_train) / 32, epochs=epochs)
 
 # here's a more "manual" example
 for e in range(epochs):
-    print 'Epoch', e
+    print('Epoch', e)
     batches = 0
-    for X_batch, Y_batch in datagen.flow(X_train, Y_train, batch_size=32):
-        loss = model.train(X_batch, Y_batch)
+    for x_batch, y_batch in datagen.flow(x_train, y_train, batch_size=32):
+        model.fit(x_batch, y_batch)
         batches += 1
-        if batches >= len(X_train) / 32:
+        if batches >= len(x_train) / 32:
             # we need to break the loop by hand because
             # the generator loops indefinitely
             break

docs/templates/preprocessing/text.md

@@ -3,7 +3,7 @@
 
 ```python
 keras.preprocessing.text.text_to_word_sequence(text, 
-    filters=base_filter(), lower=True, split=" ")
+    filters='!"#$%&()*+,-./:;<=>?@[\\]^_`{|}~\t\n', lower=True, split=" ")
 ```
 
 Split a sentence into a list of words.
@@ -12,7 +12,7 @@ Split a sentence into a list of words.
 
 - __Arguments__:
     - __text__: str.
-    - __filters__: list (or concatenation) of characters to filter out, such as punctuation. Default: base_filter(), includes basic punctuation, tabs, and newlines.
+    - __filters__: list (or concatenation) of characters to filter out, such as punctuation. Default: '!"#$%&()*+,-./:;<=>?@[\\]^_`{|}~\t\n' , includes basic punctuation, tabs, and newlines.
     - __lower__: boolean. Whether to set the text to lowercase.
     - __split__: str. Separator for word splitting.
 
@@ -20,7 +20,7 @@ Split a sentence into a list of words.
 
 ```python
 keras.preprocessing.text.one_hot(text, n,
-    filters=base_filter(), lower=True, split=" ")
+    filters='!"#$%&()*+,-./:;<=>?@[\\]^_`{|}~\t\n', lower=True, split=" ")
 ```
 
 One-hot encode a text into a list of word indexes in a vocabulary of size n.
@@ -33,14 +33,15 @@ One-hot encode a text into a list of word indexes in a vocabulary of size n.
 ## Tokenizer
 
 ```python
-keras.preprocessing.text.Tokenizer(num_words=None, filters=base_filter(), 
-    lower=True, split=" ")
+keras.preprocessing.text.Tokenizer(num_words=None, filters='!"#$%&()*+,-./:;<=>?@[\\]^_`{|}~\t\n', 
+    lower=True, split=" ", char_level=False)
 ```
 
 Class for vectorizing texts, or/and turning texts into sequences (=list of word indexes, where the word of rank i in the dataset (starting at 1) has index i).
 
 - __Arguments__: Same as `text_to_word_sequence` above.
     - __num_words__: None or int. Maximum number of words to work with (if set, tokenization will be restricted to the top num_words most common words in the dataset).
+    - __char_level__: if True, every character will be treated as a token.
 
 - __Methods__:
 

examples/addition_rnn.py

@@ -78,7 +78,7 @@ INVERT = True
 
 # Maximum length of input is 'int + int' (e.g., '345+678'). Maximum length of
 # int is DIGITS.
-MAxLEN = DIGITS + 1 + DIGITS
+MAXLEN = DIGITS + 1 + DIGITS
 
 # All the numbers, plus sign and space for padding.
 chars = '0123456789+ '
@@ -98,9 +98,9 @@ while len(questions) < TRAINING_SIZE:
     if key in seen:
         continue
     seen.add(key)
-    # Pad the data with spaces such that it is always MAxLEN.
+    # Pad the data with spaces such that it is always MAXLEN.
     q = '{}+{}'.format(a, b)
-    query = q + ' ' * (MAxLEN - len(q))
+    query = q + ' ' * (MAXLEN - len(q))
     ans = str(a + b)
     # Answers can be of maximum size DIGITS + 1.
     ans += ' ' * (DIGITS + 1 - len(ans))
@@ -113,10 +113,10 @@ while len(questions) < TRAINING_SIZE:
 print('Total addition questions:', len(questions))
 
 print('Vectorization...')
-x = np.zeros((len(questions), MAxLEN, len(chars)), dtype=np.bool)
+x = np.zeros((len(questions), MAXLEN, len(chars)), dtype=np.bool)
 y = np.zeros((len(questions), DIGITS + 1, len(chars)), dtype=np.bool)
 for i, sentence in enumerate(questions):
-    x[i] = ctable.encode(sentence, MAxLEN)
+    x[i] = ctable.encode(sentence, MAXLEN)
 for i, sentence in enumerate(expected):
     y[i] = ctable.encode(sentence, DIGITS + 1)
 
@@ -151,7 +151,7 @@ model = Sequential()
 # "Encode" the input sequence using an RNN, producing an output of HIDDEN_SIZE.
 # Note: In a situation where your input sequences have a variable length,
 # use input_shape=(None, num_feature).
-model.add(RNN(HIDDEN_SIZE, input_shape=(MAxLEN, len(chars))))
+model.add(RNN(HIDDEN_SIZE, input_shape=(MAXLEN, len(chars))))
 # As the decoder RNN's input, repeatedly provide with the last hidden state of
 # RNN for each time step. Repeat 'DIGITS + 1' times as that's the maximum
 # length of output, e.g., when DIGITS=3, max output is 999+999=1998.

examples/deep_dream.py

@@ -8,24 +8,16 @@ e.g.:
 ```
 python deep_dream.py img/mypic.jpg results/dream
 ```
-
-It is preferable to run this script on GPU, for speed.
-If running on CPU, prefer the TensorFlow backend (much faster).
-
-Example results: http://i.imgur.com/FX6ROg9.jpg
 '''
 from __future__ import print_function
 
 from keras.preprocessing.image import load_img, img_to_array
 import numpy as np
-from scipy.misc import imsave
-from scipy.optimize import fmin_l_bfgs_b
-import time
+import scipy
 import argparse
 
-from keras.applications import vgg16
+from keras.applications import inception_v3
 from keras import backend as K
-from keras.layers import Input
 
 parser = argparse.ArgumentParser(description='Deep Dreams with Keras.')
 parser.add_argument('base_image_path', metavar='base', type=str,
@@ -37,183 +29,167 @@ args = parser.parse_args()
 base_image_path = args.base_image_path
 result_prefix = args.result_prefix
 
-# dimensions of the generated picture.
-img_height = 600
-img_width = 600
-
-# some settings we found interesting
-saved_settings = {
-    'bad_trip': {'features': {'block4_conv1': 0.05,
-                              'block4_conv2': 0.01,
-                              'block4_conv3': 0.01},
-                 'continuity': 0.1,
-                 'dream_l2': 0.8,
-                 'jitter': 5},
-    'dreamy': {'features': {'block5_conv1': 0.05,
-                            'block5_conv2': 0.02},
-               'continuity': 0.1,
-               'dream_l2': 0.02,
-               'jitter': 0},
+# These are the names of the layers
+# for which we try to maximize activation,
+# as well as their weight in the final loss
+# we try to maximize.
+# You can tweak these setting to obtain new visual effects.
+settings = {
+    'features': {
+        'mixed2': 0.2,
+        'mixed3': 0.5,
+        'mixed4': 2.,
+        'mixed5': 1.5,
+    },
 }
-# the settings we will use in this experiment
-settings = saved_settings['dreamy']
 
 
 def preprocess_image(image_path):
-    # util function to open, resize and format pictures
-    # into appropriate tensors
-    img = load_img(image_path, target_size=(img_height, img_width))
+    # Util function to open, resize and format pictures
+    # into appropriate tensors.
+    img = load_img(image_path)
     img = img_to_array(img)
     img = np.expand_dims(img, axis=0)
-    img = vgg16.preprocess_input(img)
+    img = inception_v3.preprocess_input(img)
     return img
 
 
 def deprocess_image(x):
-    # util function to convert a tensor into a valid image
+    # Util function to convert a tensor into a valid image.
     if K.image_data_format() == 'channels_first':
-        x = x.reshape((3, img_height, img_width))
+        x = x.reshape((3, x.shape[2], x.shape[3]))
         x = x.transpose((1, 2, 0))
     else:
-        x = x.reshape((img_height, img_width, 3))
-    # Remove zero-center by mean pixel
-    x[:, :, 0] += 103.939
-    x[:, :, 1] += 116.779
-    x[:, :, 2] += 123.68
-    # 'BGR'->'RGB'
-    x = x[:, :, ::-1]
+        x = x.reshape((x.shape[1], x.shape[2], 3))
+    x /= 2.
+    x += 0.5
+    x *= 255.
     x = np.clip(x, 0, 255).astype('uint8')
     return x
 
-if K.image_data_format() == 'channels_first':
-    img_size = (3, img_height, img_width)
-else:
-    img_size = (img_height, img_width, 3)
-# this will contain our generated image
-dream = Input(batch_shape=(1,) + img_size)
+K.set_learning_phase(0)
 
-# build the VGG16 network with our placeholder
-# the model will be loaded with pre-trained ImageNet weights
-model = vgg16.VGG16(input_tensor=dream,
-                    weights='imagenet', include_top=False)
+# Build the InceptionV3 network with our placeholder.
+# The model will be loaded with pre-trained ImageNet weights.
+model = inception_v3.InceptionV3(weights='imagenet',
+                                 include_top=False)
+dream = model.input
 print('Model loaded.')
 
-# get the symbolic outputs of each "key" layer (we gave them unique names).
+# Get the symbolic outputs of each "key" layer (we gave them unique names).
 layer_dict = dict([(layer.name, layer) for layer in model.layers])
 
-
-def continuity_loss(x):
-    # continuity loss util function
-    assert K.ndim(x) == 4
-    if K.image_data_format() == 'channels_first':
-        a = K.square(x[:, :, :img_height - 1, :img_width - 1] -
-                     x[:, :, 1:, :img_width - 1])
-        b = K.square(x[:, :, :img_height - 1, :img_width - 1] -
-                     x[:, :, :img_height - 1, 1:])
-    else:
-        a = K.square(x[:, :img_height - 1, :img_width - 1, :] -
-                     x[:, 1:, :img_width - 1, :])
-        b = K.square(x[:, :img_height - 1, :img_width - 1, :] -
-                     x[:, :img_height - 1, 1:, :])
-    return K.sum(K.pow(a + b, 1.25))
-
-# define the loss
+# Define the loss.
 loss = K.variable(0.)
 for layer_name in settings['features']:
-    # add the L2 norm of the features of a layer to the loss
+    # Add the L2 norm of the features of a layer to the loss.
     assert layer_name in layer_dict.keys(), 'Layer ' + layer_name + ' not found in model.'
     coeff = settings['features'][layer_name]
     x = layer_dict[layer_name].output
-    shape = layer_dict[layer_name].output_shape
-    # we avoid border artifacts by only involving non-border pixels in the loss
+    # We avoid border artifacts by only involving non-border pixels in the loss.
+    scaling = K.prod(K.cast(K.shape(x), 'float32'))
     if K.image_data_format() == 'channels_first':
-        loss -= coeff * K.sum(K.square(x[:, :, 2: shape[2] - 2, 2: shape[3] - 2])) / np.prod(shape[1:])
+        loss += coeff * K.sum(K.square(x[:, :, 2: -2, 2: -2])) / scaling
     else:
-        loss -= coeff * K.sum(K.square(x[:, 2: shape[1] - 2, 2: shape[2] - 2, :])) / np.prod(shape[1:])
+        loss += coeff * K.sum(K.square(x[:, 2: -2, 2: -2, :])) / scaling
 
-# add continuity loss (gives image local coherence, can result in an artful blur)
-loss += settings['continuity'] * continuity_loss(dream) / np.prod(img_size)
-# add image L2 norm to loss (prevents pixels from taking very high values, makes image darker)
-loss += settings['dream_l2'] * K.sum(K.square(dream)) / np.prod(img_size)
+# Compute the gradients of the dream wrt the loss.
+grads = K.gradients(loss, dream)[0]
+# Normalize gradients.
+grads /= K.maximum(K.mean(K.abs(grads)), 1e-7)
 
-# feel free to further modify the loss as you see fit, to achieve new effects...
-
-# compute the gradients of the dream wrt the loss
-grads = K.gradients(loss, dream)
-
-outputs = [loss]
-if isinstance(grads, (list, tuple)):
-    outputs += grads
-else:
-    outputs.append(grads)
-
-f_outputs = K.function([dream], outputs)
+# Set up function to retrieve the value
+# of the loss and gradients given an input image.
+outputs = [loss, grads]
+fetch_loss_and_grads = K.function([dream], outputs)
 
 
 def eval_loss_and_grads(x):
-    x = x.reshape((1,) + img_size)
-    outs = f_outputs([x])
+    outs = fetch_loss_and_grads([x])
     loss_value = outs[0]
-    if len(outs[1:]) == 1:
-        grad_values = outs[1].flatten().astype('float64')
-    else:
-        grad_values = np.array(outs[1:]).flatten().astype('float64')
+    grad_values = outs[1]
     return loss_value, grad_values
 
 
-class Evaluator(object):
-    """Loss and gradients evaluator.
+def resize_img(img, size):
+    img = np.copy(img)
+    if K.image_data_format() == 'channels_first':
+        factors = (1, 1,
+                   float(size[0]) / img.shape[2],
+                   float(size[1]) / img.shape[3])
+    else:
+        factors = (1,
+                   float(size[0]) / img.shape[1],
+                   float(size[1]) / img.shape[2],
+                   1)
+    return scipy.ndimage.zoom(img, factors, order=1)
 
-    This Evaluator class makes it possible
-    to compute loss and gradients in one pass
-    while retrieving them via two separate functions,
-    "loss" and "grads". This is done because scipy.optimize
-    requires separate functions for loss and gradients,
-    but computing them separately would be inefficient.
-    """
 
-    def __init__(self):
-        self.loss_value = None
-        self.grad_values = None
-
-    def loss(self, x):
-        assert self.loss_value is None
+def gradient_ascent(x, iterations, step, max_loss=None):
+    for i in range(iterations):
         loss_value, grad_values = eval_loss_and_grads(x)
-        self.loss_value = loss_value
-        self.grad_values = grad_values
-        return self.loss_value
+        if max_loss is not None and loss_value > max_loss:
+            break
+        print('..Loss value at', i, ':', loss_value)
+        x += step * grad_values
+    return x
 
-    def grads(self, x):
-        assert self.loss_value is not None
-        grad_values = np.copy(self.grad_values)
-        self.loss_value = None
-        self.grad_values = None
-        return grad_values
 
-evaluator = Evaluator()
+def save_img(img, fname):
+    pil_img = deprocess_image(np.copy(img))
+    scipy.misc.imsave(fname, pil_img)
 
-# Run scipy-based optimization (L-BFGS) over the pixels of the generated image
-# so as to minimize the loss
-x = preprocess_image(base_image_path)
-for i in range(5):
-    print('Start of iteration', i)
-    start_time = time.time()
 
-    # Add a random jitter to the initial image.
-    # This will be reverted at decoding time
-    random_jitter = (settings['jitter'] * 2) * (np.random.random(img_size) - 0.5)
-    x += random_jitter
+"""Process:
 
-    # Run L-BFGS for 7 steps
-    x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.flatten(),
-                                     fprime=evaluator.grads, maxfun=7)
-    print('Current loss value:', min_val)
-    # Decode the dream and save it
-    x = x.reshape(img_size)
-    x -= random_jitter
-    img = deprocess_image(np.copy(x))
-    fname = result_prefix + '_at_iteration_%d.png' % i
-    imsave(fname, img)
-    end_time = time.time()
-    print('Image saved as', fname)
-    print('Iteration %d completed in %ds' % (i, end_time - start_time))
+- Load the original image.
+- Define a number of processing scales (i.e. image shapes),
+    from smallest to largest.
+- Resize the original image to the smallest scale.
+- For every scale, starting with the smallest (i.e. current one):
+    - Run gradient ascent
+    - Upscale image to the next scale
+    - Reinject the detail that was lost at upscaling time
+- Stop when we are back to the original size.
+
+To obtain the detail lost during upscaling, we simply
+take the original image, shrink it down, upscale it,
+and compare the result to the (resized) original image.
+"""
+
+
+# Playing with these hyperparameters will also allow you to achieve new effects
+step = 0.01  # Gradient ascent step size
+num_octave = 3  # Number of scales at which to run gradient ascent
+octave_scale = 1.4  # Size ratio between scales
+iterations = 20  # Number of ascent steps per scale
+max_loss = 10.
+
+img = preprocess_image(base_image_path)
+if K.image_data_format() == 'channels_first':
+    original_shape = img.shape[2:]
+else:
+    original_shape = img.shape[1:3]
+successive_shapes = [original_shape]
+for i in range(1, num_octave):
+    shape = tuple([int(dim / (octave_scale ** i)) for dim in original_shape])
+    successive_shapes.append(shape)
+successive_shapes = successive_shapes[::-1]
+original_img = np.copy(img)
+shrunk_original_img = resize_img(img, successive_shapes[0])
+
+for shape in successive_shapes:
+    print('Processing image shape', shape)
+    img = resize_img(img, shape)
+    img = gradient_ascent(img,
+                          iterations=iterations,
+                          step=step,
+                          max_loss=max_loss)
+    upscaled_shrunk_original_img = resize_img(shrunk_original_img, shape)
+    same_size_original = resize_img(original_img, shape)
+    lost_detail = same_size_original - upscaled_shrunk_original_img
+
+    img += lost_detail
+    shrunk_original_img = resize_img(original_img, shape)
+
+save_img(img, fname=result_prefix + '.png')

examples/imdb_bidirectional_lstm.py

@@ -24,7 +24,7 @@ print('Loading data...')
 print(len(x_train), 'train sequences')
 print(len(x_test), 'test sequences')
 
-print("Pad sequences (samples x time)")
+print('Pad sequences (samples x time)')
 x_train = sequence.pad_sequences(x_train, maxlen=maxlen)
 x_test = sequence.pad_sequences(x_test, maxlen=maxlen)
 print('x_train shape:', x_train.shape)

examples/mnist_siamese_graph.py

@@ -24,7 +24,7 @@ from keras import backend as K
 
 def euclidean_distance(vects):
     x, y = vects
-    return K.sqrt(K.sum(K.square(x - y), axis=1, keepdims=True))
+    return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon()))
 
 
 def eucl_dist_output_shape(shapes):

examples/neural_doodle.py

@@ -196,8 +196,8 @@ x = mask_input
 for layer in image_model.layers[1:]:
     name = 'mask_%s' % layer.name
     if 'conv' in layer.name:
-        x = AveragePooling2D((3, 3), strides=(
-            1, 1), name=name, border_mode='same')(x)
+        x = AveragePooling2D((3, 3), padding='same', strides=(
+            1, 1), name=name)(x)
     elif 'pool' in layer.name:
         x = AveragePooling2D((2, 2), name=name)(x)
 mask_model = Model(mask_input, x)
@@ -238,6 +238,7 @@ def region_style_loss(style_image, target_image, style_mask, target_mask):
         masked_target = K.permute_dimensions(
             target_image, (2, 0, 1)) * target_mask
         num_channels = K.shape(style_image)[-1]
+    num_channels = K.cast(num_channels, dtype='float32')
     s = gram_matrix(masked_style) / K.mean(style_mask) / num_channels
     c = gram_matrix(masked_target) / K.mean(target_mask) / num_channels
     return K.mean(K.square(s - c))

examples/neural_style_transfer.py

@@ -57,7 +57,7 @@ from scipy.optimize import fmin_l_bfgs_b
 import time
 import argparse
 
-from keras.applications import vgg16
+from keras.applications import vgg19
 from keras import backend as K
 
 parser = argparse.ArgumentParser(description='Neural style transfer with Keras.')
@@ -99,7 +99,7 @@ def preprocess_image(image_path):
     img = load_img(image_path, target_size=(img_nrows, img_ncols))
     img = img_to_array(img)
     img = np.expand_dims(img, axis=0)
-    img = vgg16.preprocess_input(img)
+    img = vgg19.preprocess_input(img)
     return img
 
 # util function to convert a tensor into a valid image
@@ -137,7 +137,7 @@ input_tensor = K.concatenate([base_image,
 
 # build the VGG16 network with our 3 images as input
 # the model will be loaded with pre-trained ImageNet weights
-model = vgg16.VGG16(input_tensor=input_tensor,
+model = vgg19.VGG19(input_tensor=input_tensor,
                     weights='imagenet', include_top=False)
 print('Model loaded.')
 
@@ -199,7 +199,7 @@ def total_variation_loss(x):
 
 # combine these loss functions into a single scalar
 loss = K.variable(0.)
-layer_features = outputs_dict['block4_conv2']
+layer_features = outputs_dict['block5_conv2']
 base_image_features = layer_features[0, :, :, :]
 combination_features = layer_features[2, :, :, :]
 loss += content_weight * content_loss(base_image_features,
@@ -273,10 +273,7 @@ evaluator = Evaluator()
 
 # run scipy-based optimization (L-BFGS) over the pixels of the generated image
 # so as to minimize the neural style loss
-if K.image_data_format() == 'channels_first':
-    x = np.random.uniform(0, 255, (1, 3, img_nrows, img_ncols)) - 128.
-else:
-    x = np.random.uniform(0, 255, (1, img_nrows, img_ncols, 3)) - 128.
+x = preprocess_image(base_image_path)
 
 for i in range(iterations):
     print('Start of iteration', i)

examples/variational_autoencoder.py

@@ -6,7 +6,7 @@ import numpy as np
 import matplotlib.pyplot as plt
 from scipy.stats import norm
 
-from keras.layers import Input, Dense, Lambda
+from keras.layers import Input, Dense, Lambda, Layer
 from keras.models import Model
 from keras import backend as K
 from keras import metrics
@@ -19,6 +19,7 @@ intermediate_dim = 256
 epochs = 50
 epsilon_std = 1.0
 
+
 x = Input(batch_shape=(batch_size, original_dim))
 h = Dense(intermediate_dim, activation='relu')(x)
 z_mean = Dense(latent_dim)(h)
@@ -41,13 +42,29 @@ h_decoded = decoder_h(z)
 x_decoded_mean = decoder_mean(h_decoded)
 
 
-def vae_loss(x, x_decoded_mean):
-    xent_loss = original_dim * metrics.binary_crossentropy(x, x_decoded_mean)
-    kl_loss = - 0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
-    return xent_loss + kl_loss
+# Custom loss layer
+class CustomVariationalLayer(Layer):
+    def __init__(self, **kwargs):
+        self.is_placeholder = True
+        super(CustomVariationalLayer, self).__init__(**kwargs)
+
+    def vae_loss(self, x, x_decoded_mean):
+        xent_loss = original_dim * metrics.binary_crossentropy(x, x_decoded_mean)
+        kl_loss = - 0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
+        return K.mean(xent_loss + kl_loss)
+
+    def call(self, inputs):
+        x = inputs[0]
+        x_decoded_mean = inputs[1]
+        loss = self.vae_loss(x, x_decoded_mean)
+        self.add_loss(loss, inputs=inputs)
+        # We won't actually use the output.
+        return x
+
+y = CustomVariationalLayer()([x, x_decoded_mean])
+vae = Model(x, y)
+vae.compile(optimizer='rmsprop', loss=None)
 
-vae = Model(x, x_decoded_mean)
-vae.compile(optimizer='rmsprop', loss=vae_loss)
 
 # train the VAE on MNIST digits
 (x_train, y_train), (x_test, y_test) = mnist.load_data()
@@ -57,7 +74,7 @@ x_test = x_test.astype('float32') / 255.
 x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:])))
 x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))
 
-vae.fit(x_train, x_train,
+vae.fit(x_train,
         shuffle=True,
         epochs=epochs,
         batch_size=batch_size,

examples/variational_autoencoder_deconv.py

@@ -7,7 +7,7 @@ import numpy as np
 import matplotlib.pyplot as plt
 from scipy.stats import norm
 
-from keras.layers import Input, Dense, Lambda, Flatten, Reshape
+from keras.layers import Input, Dense, Lambda, Flatten, Reshape, Layer
 from keras.layers import Conv2D, Conv2DTranspose
 from keras.models import Model
 from keras import backend as K
@@ -106,17 +106,31 @@ x_decoded_relu = decoder_deconv_3_upsamp(deconv_2_decoded)
 x_decoded_mean_squash = decoder_mean_squash(x_decoded_relu)
 
 
-def vae_loss(x, x_decoded_mean):
-    # NOTE: binary_crossentropy expects a batch_size by dim
-    # for x and x_decoded_mean, so we MUST flatten these!
-    x = K.flatten(x)
-    x_decoded_mean = K.flatten(x_decoded_mean)
-    xent_loss = img_rows * img_cols * metrics.binary_crossentropy(x, x_decoded_mean)
-    kl_loss = - 0.5 * K.mean(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
-    return xent_loss + kl_loss
+# Custom loss layer
+class CustomVariationalLayer(Layer):
+    def __init__(self, **kwargs):
+        self.is_placeholder = True
+        super(CustomVariationalLayer, self).__init__(**kwargs)
 
-vae = Model(x, x_decoded_mean_squash)
-vae.compile(optimizer='rmsprop', loss=vae_loss)
+    def vae_loss(self, x, x_decoded_mean_squash):
+        x = K.flatten(x)
+        x_decoded_mean_squash = K.flatten(x_decoded_mean_squash)
+        xent_loss = img_rows * img_cols * metrics.binary_crossentropy(x, x_decoded_mean_squash)
+        kl_loss = - 0.5 * K.mean(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
+        return K.mean(xent_loss + kl_loss)
+
+    def call(self, inputs):
+        x = inputs[0]
+        x_decoded_mean_squash = inputs[1]
+        loss = self.vae_loss(x, x_decoded_mean_squash)
+        self.add_loss(loss, inputs=inputs)
+        # We don't use this output.
+        return x
+
+
+y = CustomVariationalLayer()([x, x_decoded_mean_squash])
+vae = Model(x, y)
+vae.compile(optimizer='rmsprop', loss=None)
 vae.summary()
 
 # train the VAE on MNIST digits
@@ -129,7 +143,7 @@ x_test = x_test.reshape((x_test.shape[0],) + original_img_size)
 
 print('x_train.shape:', x_train.shape)
 
-vae.fit(x_train, x_train,
+vae.fit(x_train,
         shuffle=True,
         epochs=epochs,
         batch_size=batch_size,

keras/__init__.py

@@ -17,5 +17,7 @@ from . import models
 from . import losses
 from . import optimizers
 from . import regularizers
+# Importable from root because it's technically not a layer
+from .layers import Input
 
-__version__ = '2.0.3'
+__version__ = '2.0.5'

keras/activations.py

@@ -1,7 +1,9 @@
 from __future__ import absolute_import
 import six
+import warnings
 from . import backend as K
 from .utils.generic_utils import deserialize_keras_object
+from .engine import Layer
 
 
 def softmax(x, axis=-1):
@@ -78,6 +80,13 @@ def get(identifier):
         identifier = str(identifier)
         return deserialize(identifier)
     elif callable(identifier):
+        if isinstance(identifier, Layer):
+            warnings.warn((
+                'Do not pass a layer instance (such as {identifier}) as the '
+                'activation argument of another layer. Instead, advanced '
+                'activation layers should be used just like any other '
+                'layer in a model.'
+            ).format(identifier=identifier.__class__.__name__))
         return identifier
     else:
         raise ValueError('Could not interpret '

keras/applications/inception_v3.py

@@ -27,7 +27,6 @@ from ..layers import AveragePooling2D
 from ..layers import GlobalAveragePooling2D
 from ..layers import GlobalMaxPooling2D
 from ..engine.topology import get_source_inputs
-from ..utils.layer_utils import convert_all_kernels_in_model
 from ..utils.data_utils import get_file
 from .. import backend as K
 from .imagenet_utils import decode_predictions
@@ -384,8 +383,6 @@ def InceptionV3(include_top=True,
                 cache_subdir='models',
                 md5_hash='bcbd6486424b2319ff4ef7d526e38f63')
         model.load_weights(weights_path)
-        if K.backend() == 'theano':
-            convert_all_kernels_in_model(model)
     return model
 
 

keras/applications/resnet50.py

@@ -43,7 +43,7 @@ def identity_block(input_tensor, kernel_size, filters, stage, block):
 
     # Arguments
         input_tensor: input tensor
-        kernel_size: defualt 3, the kernel size of middle conv layer at main path
+        kernel_size: default 3, the kernel size of middle conv layer at main path
         filters: list of integers, the filterss of 3 conv layer at main path
         stage: integer, current stage label, used for generating layer names
         block: 'a','b'..., current block label, used for generating layer names
@@ -81,7 +81,7 @@ def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2))
 
     # Arguments
         input_tensor: input tensor
-        kernel_size: defualt 3, the kernel size of middle conv layer at main path
+        kernel_size: default 3, the kernel size of middle conv layer at main path
         filters: list of integers, the filterss of 3 conv layer at main path
         stage: integer, current stage label, used for generating layer names
         block: 'a','b'..., current block label, used for generating layer names
@@ -264,21 +264,19 @@ def ResNet50(include_top=True, weights='imagenet',
         model.load_weights(weights_path)
         if K.backend() == 'theano':
             layer_utils.convert_all_kernels_in_model(model)
-
-        if K.image_data_format() == 'channels_first':
             if include_top:
                 maxpool = model.get_layer(name='avg_pool')
                 shape = maxpool.output_shape[1:]
                 dense = model.get_layer(name='fc1000')
                 layer_utils.convert_dense_weights_data_format(dense, shape, 'channels_first')
 
-            if K.backend() == 'tensorflow':
-                warnings.warn('You are using the TensorFlow backend, yet you '
-                              'are using the Theano '
-                              'image data format convention '
-                              '(`image_data_format="channels_first"`). '
-                              'For best performance, set '
-                              '`image_data_format="channels_last"` in '
-                              'your Keras config '
-                              'at ~/.keras/keras.json.')
+        if K.image_data_format() == 'channels_first' and K.backend() == 'tensorflow':
+            warnings.warn('You are using the TensorFlow backend, yet you '
+                          'are using the Theano '
+                          'image data format convention '
+                          '(`image_data_format="channels_first"`). '
+                          'For best performance, set '
+                          '`image_data_format="channels_last"` in '
+                          'your Keras config '
+                          'at ~/.keras/keras.json.')
     return model

keras/applications/vgg16.py

@@ -59,7 +59,7 @@ def VGG16(include_top=True, weights='imagenet',
         input_shape: optional shape tuple, only to be specified
             if `include_top` is False (otherwise the input shape
             has to be `(224, 224, 3)` (with `channels_last` data format)
-            or `(3, 224, 244)` (with `channels_first` data format).
+            or `(3, 224, 224)` (with `channels_first` data format).
             It should have exactly 3 inputs channels,
             and width and height should be no smaller than 48.
             E.g. `(200, 200, 3)` would be one valid value.

keras/applications/vgg19.py

@@ -59,7 +59,7 @@ def VGG19(include_top=True, weights='imagenet',
         input_shape: optional shape tuple, only to be specified
             if `include_top` is False (otherwise the input shape
             has to be `(224, 224, 3)` (with `channels_last` data format)
-            or `(3, 224, 244)` (with `channels_first` data format).
+            or `(3, 224, 224)` (with `channels_first` data format).
             It should have exactly 3 inputs channels,
             and width and height should be no smaller than 48.
             E.g. `(200, 200, 3)` would be one valid value.

keras/backend/__init__.py

@@ -10,7 +10,6 @@ from .common import set_floatx
 from .common import cast_to_floatx
 from .common import image_data_format
 from .common import set_image_data_format
-from .common import is_keras_tensor
 
 # Obtain Keras base dir path: either ~/.keras or /tmp.
 _keras_base_dir = os.path.expanduser('~')
@@ -33,7 +32,7 @@ if os.path.exists(_config_path):
     _epsilon = _config.get('epsilon', epsilon())
     assert isinstance(_epsilon, float)
     _backend = _config.get('backend', _BACKEND)
-    assert _backend in {'theano', 'tensorflow'}
+    assert _backend in {'theano', 'tensorflow', 'cntk'}
     _image_data_format = _config.get('image_data_format',
                                      image_data_format())
     assert _image_data_format in {'channels_last', 'channels_first'}
@@ -44,28 +43,39 @@ if os.path.exists(_config_path):
     _BACKEND = _backend
 
 # Save config file, if possible.
-if os.access(_keras_base_dir, os.W_OK):
-    if not os.path.exists(_keras_dir):
-        try:
-            os.makedirs(_keras_dir)
-        except OSError:
-            pass
-    if not os.path.exists(_config_path):
-        _config = {'floatx': floatx(),
-                   'epsilon': epsilon(),
-                   'backend': _BACKEND,
-                   'image_data_format': image_data_format()}
+if not os.path.exists(_keras_dir):
+    try:
+        os.makedirs(_keras_dir)
+    except OSError:
+        # Except permission denied and potential race conditions
+        # in multi-threaded environments.
+        pass
+
+if not os.path.exists(_config_path):
+    _config = {
+        'floatx': floatx(),
+        'epsilon': epsilon(),
+        'backend': _BACKEND,
+        'image_data_format': image_data_format()
+    }
+    try:
         with open(_config_path, 'w') as f:
             f.write(json.dumps(_config, indent=4))
+    except IOError:
+        # Except permission denied.
+        pass
 
 # Set backend based on KERAS_BACKEND flag, if applicable.
 if 'KERAS_BACKEND' in os.environ:
     _backend = os.environ['KERAS_BACKEND']
-    assert _backend in {'theano', 'tensorflow'}
+    assert _backend in {'theano', 'tensorflow', 'cntk'}
     _BACKEND = _backend
 
 # Import backend functions.
-if _BACKEND == 'theano':
+if _BACKEND == 'cntk':
+    sys.stderr.write('Using CNTK backend\n')
+    from .cntk_backend import *
+elif _BACKEND == 'theano':
     sys.stderr.write('Using Theano backend.\n')
     from .theano_backend import *
 elif _BACKEND == 'tensorflow':

keras/backend/common.py

@@ -44,7 +44,7 @@ def set_epsilon(e):
 
 
 def floatx():
-    """Returns the default float type, as a string
+    """Returns the default float type, as a string.
     (e.g. 'float16', 'float32', 'float64').
 
     # Returns
@@ -109,8 +109,7 @@ def cast_to_floatx(x):
 
 
 def image_data_format():
-    """Returns the default image data format
-    convention ('channels_first' or 'channels_last').
+    """Returns the default image data format convention ('channels_first' or 'channels_last').
 
     # Returns
         A string, either `'channels_first'` or `'channels_last'`
@@ -146,42 +145,13 @@ def set_image_data_format(data_format):
     _IMAGE_DATA_FORMAT = str(data_format)
 
 
-def is_keras_tensor(x):
-    """Returns whether `x` is a Keras tensor.
-
-    # Arguments
-        x: a potential tensor.
-
-    # Returns
-        A boolean: whether the argument is a Keras tensor.
-
-    # Examples
-    ```python
-        >>> from keras import backend as K
-        >>> np_var = numpy.array([1, 2])
-        >>> K.is_keras_tensor(np_var)
-        False
-        >>> keras_var = K.variable(np_var)
-        >>> K.is_keras_tensor(keras_var)  # A variable is not a Tensor.
-        False
-        >>> keras_placeholder = K.placeholder(shape=(2, 4, 5))
-        >>> K.is_keras_tensor(keras_placeholder)  # A placeholder is a Tensor.
-        True
-    ```
-    """
-    if hasattr(x, '_keras_shape'):
-        return True
-    else:
-        return False
-
-
 # Legacy methods
 
 def set_image_dim_ordering(dim_ordering):
     """Legacy setter for `image_data_format`.
 
     # Arguments
-        dim_ordering: string. `'tf'` or `'th'`.
+        dim_ordering: string. `tf` or `th`.
 
     # Example
     ```python
@@ -192,6 +162,9 @@ def set_image_dim_ordering(dim_ordering):
         >>> K.image_data_format()
         'channels_last'
     ```
+
+    # Raises
+        ValueError: if `dim_ordering` is invalid.
     """
     global _IMAGE_DATA_FORMAT
     if dim_ordering not in {'tf', 'th'}:
@@ -205,6 +178,9 @@ def set_image_dim_ordering(dim_ordering):
 
 def image_dim_ordering():
     """Legacy getter for `image_data_format`.
+
+    # Returns
+        string, one of `'th'`, `'tf'`
     """
     if _IMAGE_DATA_FORMAT == 'channels_first':
         return 'th'

keras/backend/tensorflow_backend.py

@@ -4,17 +4,20 @@ from tensorflow.python.ops import tensor_array_ops
 from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import functional_ops
 from tensorflow.python.ops import ctc_ops as ctc
+from tensorflow.python.ops import variables as tf_variables
 
 from collections import defaultdict
+import inspect
 import numpy as np
 import os
-import warnings
 
 from .common import floatx
 from .common import _EPSILON
 from .common import image_data_format
+
 # Legacy functions
-from .common import set_image_dim_ordering, image_dim_ordering
+from .common import set_image_dim_ordering
+from .common import image_dim_ordering
 
 py_all = all
 py_sum = sum
@@ -43,6 +46,14 @@ _MANUAL_VAR_INIT = False
 
 
 def get_uid(prefix=''):
+    """Get the uid for the default graph.
+
+    # Arguments
+        prefix: An optional prefix of the graph.
+
+    # Returns
+        A unique identifier for the graph.
+    """
     global _GRAPH_UID_DICTS
     graph = tf.get_default_graph()
     if graph not in _GRAPH_UID_DICTS:
@@ -52,6 +63,7 @@ def get_uid(prefix=''):
 
 
 def reset_uids():
+    """Reset graph identifiers."""
     global _GRAPH_UID_DICTS
     _GRAPH_UID_DICTS = {}
 
@@ -169,6 +181,17 @@ def set_session(session):
 # VARIABLE MANIPULATION
 
 def _convert_string_dtype(dtype):
+    """Get the type from a string.
+
+    # Arguments
+        dtype: A string representation of a type.
+
+    # Returns
+        The type requested.
+
+    # Raises
+        ValueError: if `dtype` is not supported.
+    """
     if dtype == 'float16':
         return tf.float16
     if dtype == 'float32':
@@ -190,6 +213,15 @@ def _convert_string_dtype(dtype):
 
 
 def _to_tensor(x, dtype):
+    """Convert the input `x` to a tensor of type `dtype`.
+
+    # Arguments
+        x: An object to be converted (numpy array, list, tensors).
+        dtype: The destination type.
+
+    # Returns
+        A tensor.
+    """
     x = tf.convert_to_tensor(x)
     if x.dtype != dtype:
         x = tf.cast(x, dtype)
@@ -309,11 +341,59 @@ def _initialize_variables():
 
 
 def constant(value, dtype=None, shape=None, name=None):
+    """Creates a constant tensor.
+
+    # Arguments
+        value: A constant value (or list)
+        dtype: The type of the elements of the resulting tensor.
+        shape: Optional dimensions of resulting tensor.
+        name: Optional name for the tensor.
+
+    # Returns
+        A Constant Tensor.
+    """
     if dtype is None:
         dtype = floatx()
     return tf.constant(value, dtype=dtype, shape=shape, name=name)
 
 
+def is_keras_tensor(x):
+    """Returns whether `x` is a Keras tensor.
+
+    # Arguments
+        x: a potential tensor.
+
+    # Returns
+        A boolean: whether the argument is a Keras tensor.
+
+    # Raises
+        ValueError: in case `x` is not a symbolic tensor.
+
+    # Examples
+    ```python
+        >>> from keras import backend as K
+        >>> np_var = numpy.array([1, 2])
+        >>> K.is_keras_tensor(np_var) # A numpy array is not a symbolic yensor.
+        ValueError
+        >>> k_var = tf.placeholder('float32', shape=(1,1))
+        >>> K.is_keras_tensor(k_var) # A variable created directly from tensorflow/theano is not a Keras tensor.
+        False
+        >>> keras_var = K.variable(np_var)
+        >>> K.is_keras_tensor(keras_var)  # A variable created with the keras backend is a Keras tensor.
+        True
+        >>> keras_placeholder = K.placeholder(shape=(2, 4, 5))
+        >>> K.is_keras_tensor(keras_placeholder)  # A placeholder is a Keras tensor.
+        True
+    ```
+    """
+    if not isinstance(x, (tf.Tensor,
+                          tf_variables.Variable,
+                          tf.SparseTensor)):
+        raise ValueError('Unexpectedly found an instance of type `' + str(type(x)) + '`. '
+                         'Expected a symbolic tensor instance.')
+    return hasattr(x, '_keras_history')
+
+
 def placeholder(shape=None, ndim=None, dtype=None, sparse=False, name=None):
     """Instantiates a placeholder tensor and returns it.
 
@@ -626,6 +706,18 @@ def ones_like(x, dtype=None, name=None):
     return tf.ones_like(x, dtype=dtype, name=name)
 
 
+def identity(x):
+    """Returns a tensor with the same content as the input tensor.
+
+    # Arguments
+        x: The input tensor.
+
+    # Returns
+        A tensor of the same shape, type and content.
+    """
+    return tf.identity(x)
+
+
 def random_uniform_variable(shape, low, high, dtype=None,
                             name=None, seed=None):
     """Instantiates a variable with values drawn from a uniform distribution.
@@ -761,18 +853,54 @@ def cast(x, dtype):
 
 
 def update(x, new_x):
+    """Update the value of `x` to `new_x`.
+
+    # Arguments
+        x: A Variable.
+        new_x: A tensor of same shape as `x`.
+
+    # Returns
+        The variable `x` updated.
+    """
     return tf.assign(x, new_x)
 
 
 def update_add(x, increment):
+    """Update the value of `x` by adding `increment`.
+
+    # Arguments
+        x: A Variable.
+        increment: A tensor of same shape as `x`.
+
+    # Returns
+        The variable `x` updated.
+    """
     return tf.assign_add(x, increment)
 
 
 def update_sub(x, decrement):
+    """Update the value of `x` by subtracting `decrement`.
+
+    # Arguments
+        x: A Variable.
+        decrement: A tensor of same shape as `x`.
+
+    # Returns
+        The variable `x` updated.
+    """
     return tf.assign_sub(x, decrement)
 
 
 def moving_average_update(x, value, momentum):
+    """Compute the moving average of a variable.
+
+    # Arguments
+        x: A Variable.
+        value: A tensor with the same shape as `variable`.
+        momentum: The moving average momentum.
+
+    # Returns
+        An Operation to update the variable."""
     return moving_averages.assign_moving_average(
         x, value, momentum, zero_debias=False)
 
@@ -902,6 +1030,16 @@ def batch_dot(x, y, axes=None):
     """
     if isinstance(axes, int):
         axes = (axes, axes)
+    x_ndim = ndim(x)
+    y_ndim = ndim(y)
+    if x_ndim > y_ndim:
+        diff = x_ndim - y_ndim
+        y = tf.reshape(y, tf.concat([tf.shape(y), [1] * (diff)], axis=0))
+    elif y_ndim > x_ndim:
+        diff = y_ndim - x_ndim
+        x = tf.reshape(x, tf.concat([tf.shape(x), [1] * (diff)], axis=0))
+    else:
+        diff = 0
     if ndim(x) == 2 and ndim(y) == 2:
         if axes[0] == axes[1]:
             out = tf.reduce_sum(tf.multiply(x, y), axes[0])
@@ -915,6 +1053,12 @@ def batch_dot(x, y, axes=None):
             adj_x = None
             adj_y = None
         out = tf.matmul(x, y, adjoint_a=adj_x, adjoint_b=adj_y)
+    if diff:
+        if x_ndim > y_ndim:
+            idx = x_ndim + y_ndim - 3
+        else:
+            idx = x_ndim - 1
+        out = tf.squeeze(out, list(range(idx, idx + diff)))
     if ndim(out) == 1:
         out = expand_dims(out, 1)
     return out
@@ -1276,6 +1420,28 @@ def log(x):
     return tf.log(x)
 
 
+def logsumexp(x, axis=None, keepdims=False):
+    """Computes log(sum(exp(elements across dimensions of a tensor))).
+
+    This function is more numerically stable than log(sum(exp(x))).
+    It avoids overflows caused by taking the exp of large inputs and
+    underflows caused by taking the log of small inputs.
+
+    # Arguments
+        x: A tensor or variable.
+        axis: An integer, the axis to reduce over.
+        keepdims: A boolean, whether to keep the dimensions or not.
+            If `keepdims` is `False`, the rank of the tensor is reduced
+            by 1. If `keepdims` is `True`, the reduced dimension is
+            retained with length 1.
+
+    # Returns
+        The reduced tensor.
+    """
+    axis = _normalize_axis(axis, ndim(x))
+    return tf.reduce_logsumexp(x, axis=axis, keep_dims=keepdims)
+
+
 def round(x):
     """Element-wise rounding to the closest integer.
 
@@ -1589,14 +1755,13 @@ def resize_images(x, height_factor, width_factor, data_format):
         x: Tensor or variable to resize.
         height_factor: Positive integer.
         width_factor: Positive integer.
-        data_format: One of `"channels_first"`, `"channels_last"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
 
     # Returns
         A tensor.
 
     # Raises
-        ValueError: if `data_format` is neither
-            `channels_last` or `channels_first`.
+        ValueError: if `data_format` is neither `"channels_last"` or `"channels_first"`.
     """
     if data_format == 'channels_first':
         original_shape = int_shape(x)
@@ -1628,14 +1793,13 @@ def resize_volumes(x, depth_factor, height_factor, width_factor, data_format):
         depth_factor: Positive integer.
         height_factor: Positive integer.
         width_factor: Positive integer.
-        data_format: One of `"channels_first"`, `"channels_last"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
 
     # Returns
         A tensor.
 
     # Raises
-        ValueError: if `data_format` is neither
-            `channels_last` or `channels_first`.
+        ValueError: if `data_format` is neither `"channels_last"` or `"channels_first"`.
     """
     if data_format == 'channels_first':
         output = repeat_elements(x, depth_factor, axis=2)
@@ -1774,14 +1938,14 @@ def batch_flatten(x):
 
 
 def expand_dims(x, axis=-1):
-    """Adds a 1-sized dimension at index "dim".
+    """Adds a 1-sized dimension at index "axis".
 
     # Arguments
         x: A tensor or variable.
         axis: Position where to add a new axis.
 
     # Returns
-        A tensor with expended dimensions.
+        A tensor with expanded dimensions.
     """
     return tf.expand_dims(x, axis)
 
@@ -1821,14 +1985,13 @@ def spatial_2d_padding(x, padding=((1, 1), (1, 1)), data_format=None):
     # Arguments
         x: Tensor or variable.
         padding: Tuple of 2 tuples, padding pattern.
-        data_format: One of `channels_last` or `channels_first`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
 
     # Returns
         A padded 4D tensor.
 
     # Raises
-        ValueError: if `data_format` is neither
-            `channels_last` or `channels_first`.
+        ValueError: if `data_format` is neither `"channels_last"` or `"channels_first"`.
     """
     assert len(padding) == 2
     assert len(padding[0]) == 2
@@ -1864,14 +2027,13 @@ def spatial_3d_padding(x, padding=((1, 1), (1, 1), (1, 1)), data_format=None):
     # Arguments
         x: Tensor or variable.
         padding: Tuple of 3 tuples, padding pattern.
-        data_format: One of `channels_last` or `channels_first`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
 
     # Returns
         A padded 5D tensor.
 
     # Raises
-        ValueError: if `data_format` is neither
-            `channels_last` or `channels_first`.
+        ValueError: if `data_format` is neither `"channels_last"` or `"channels_first"`.
 
     """
     assert len(padding) == 3
@@ -2061,9 +2223,10 @@ class Function(object):
         inputs: Feed placeholders to the computation graph.
         outputs: Output tensors to fetch.
         updates: Additional update ops to be run at function call.
+        name: a name to help users identify what this function does.
     """
 
-    def __init__(self, inputs, outputs, updates=None):
+    def __init__(self, inputs, outputs, updates=None, name=None, **session_kwargs):
         updates = updates or []
         if not isinstance(inputs, (list, tuple)):
             raise TypeError('`inputs` to a TensorFlow backend function '
@@ -2086,6 +2249,8 @@ class Function(object):
                     # assumed already an op
                     updates_ops.append(update)
             self.updates_op = tf.group(*updates_ops)
+        self.name = name
+        self.session_kwargs = session_kwargs
 
     def __call__(self, inputs):
         if not isinstance(inputs, (list, tuple)):
@@ -2100,7 +2265,8 @@ class Function(object):
             feed_dict[tensor] = value
         session = get_session()
         updated = session.run(self.outputs + [self.updates_op],
-                              feed_dict=feed_dict)
+                              feed_dict=feed_dict,
+                              **self.session_kwargs)
         return updated[:len(self.outputs)]
 
 
@@ -2111,18 +2277,21 @@ def function(inputs, outputs, updates=None, **kwargs):
         inputs: List of placeholder tensors.
         outputs: List of output tensors.
         updates: List of update ops.
-        **kwargs: Not used with TensorFlow.
+        **kwargs: Passed to `tf.Session.run`.
 
     # Returns
         Output values as Numpy arrays.
+
+    # Raises
+        ValueError: if invalid kwargs are passed in.
     """
     if kwargs:
-        msg = [
-            'Expected no kwargs, you passed %s' % len(kwargs),
-            'kwargs passed to function are ignored with Tensorflow backend'
-        ]
-        warnings.warn('\n'.join(msg))
-    return Function(inputs, outputs, updates=updates)
+        for key in kwargs:
+            if (key not in inspect.getargspec(tf.Session.run)[0] and
+                    key not in inspect.getargspec(Function.__init__)[0]):
+                msg = 'Invalid argument "%s" passed to K.function with Tensorflow backend' % key
+                raise ValueError(msg)
+    return Function(inputs, outputs, updates=updates, **kwargs)
 
 
 def gradients(loss, variables):
@@ -2600,7 +2769,7 @@ def sparse_categorical_crossentropy(output, target, from_logits=False):
     # Returns
         Output tensor.
     """
-    # Note: tf.nn.softmax_cross_entropy_with_logits
+    # Note: tf.nn.sparse_softmax_cross_entropy_with_logits
     # expects logits, Keras expects probabilities.
     if not from_logits:
         epsilon = _to_tensor(_EPSILON, output.dtype.base_dtype)
@@ -2633,7 +2802,7 @@ def binary_crossentropy(output, target, from_logits=False):
     # Returns
         A tensor.
     """
-    # Note: tf.nn.softmax_cross_entropy_with_logits
+    # Note: tf.nn.sigmoid_cross_entropy_with_logits
     # expects logits, Keras expects probabilities.
     if not from_logits:
         # transform back to logits
@@ -2707,7 +2876,7 @@ def dropout(x, level, noise_shape=None, seed=None):
     if seed is None:
         seed = np.random.randint(10e6)
     # the dummy 1. works around a TF bug
-    # (float32_ref vs. float32 incomptability)
+    # (float32_ref vs. float32 incompatibility)
     return tf.nn.dropout(x * 1., retain_prob, noise_shape, seed=seed)
 
 
@@ -2745,6 +2914,16 @@ def in_top_k(predictions, targets, k):
 # CONVOLUTIONS
 
 def _preprocess_deconv_output_shape(x, shape, data_format):
+    """Get the output_shape for the deconvolution.
+
+    # Arguments
+        x: input tensor.
+        shape: output shape.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+
+    # Returns
+        The output shape.
+    """
     if data_format == 'channels_first':
         shape = (shape[0], shape[2], shape[3], shape[1])
 
@@ -2755,6 +2934,15 @@ def _preprocess_deconv_output_shape(x, shape, data_format):
 
 
 def _preprocess_conv2d_input(x, data_format):
+    """Transpose and cast the input before the conv2d.
+
+    # Arguments
+        x: input tensor.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+
+    # Returns
+        A tensor.
+    """
     if dtype(x) == 'float64':
         x = tf.cast(x, 'float32')
     if data_format == 'channels_first':
@@ -2767,6 +2955,15 @@ def _preprocess_conv2d_input(x, data_format):
 
 
 def _preprocess_conv3d_input(x, data_format):
+    """Transpose and cast the input before the conv3d.
+
+    # Arguments
+        x: input tensor.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+
+    # Returns
+        A tensor.
+    """
     if dtype(x) == 'float64':
         x = tf.cast(x, 'float32')
     if data_format == 'channels_first':
@@ -2775,6 +2972,15 @@ def _preprocess_conv3d_input(x, data_format):
 
 
 def _preprocess_conv2d_kernel(kernel, data_format):
+    """Transpose and cast the kernel before the conv2d.
+
+    # Arguments
+        kernel: kernel tensor.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+
+    # Returns
+        A tensor.
+    """
     if dtype(kernel) == 'float64':
         kernel = tf.cast(kernel, 'float32')
     if data_format == 'channels_first':
@@ -2783,6 +2989,15 @@ def _preprocess_conv2d_kernel(kernel, data_format):
 
 
 def _preprocess_conv3d_kernel(kernel, data_format):
+    """Transpose and cast the kernel before the conv3d.
+
+    # Arguments
+        kernel: kernel tensor.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+
+    # Returns
+        A tensor.
+    """
     if dtype(kernel) == 'float64':
         kernel = tf.cast(kernel, 'float32')
     if data_format == 'channels_first':
@@ -2791,16 +3006,37 @@ def _preprocess_conv3d_kernel(kernel, data_format):
 
 
 def _preprocess_padding(padding):
+    """Convert keras' padding to tensorflow's padding.
+
+    # Arguments
+        padding: string, `"same"` or `"valid"`.
+
+    # Returns
+        a string, `"SAME"` or `"VALID"`.
+
+    # Raises
+        ValueError: if `padding` is invalid.
+    """
     if padding == 'same':
         padding = 'SAME'
     elif padding == 'valid':
         padding = 'VALID'
     else:
-        raise ValueError('Invalid border mode:', padding)
+        raise ValueError('Invalid padding:', padding)
     return padding
 
 
 def _postprocess_conv2d_output(x, data_format):
+    """Transpose and cast the output from conv2d if needed.
+
+    # Arguments
+        x: A tensor.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+
+    # Returns
+        A tensor.
+    """
+
     if data_format == 'channels_first':
         x = tf.transpose(x, (0, 3, 1, 2))
 
@@ -2810,6 +3046,15 @@ def _postprocess_conv2d_output(x, data_format):
 
 
 def _postprocess_conv3d_output(x, data_format):
+    """Transpose and cast the output from conv3d if needed.
+
+    # Arguments
+        x: A tensor.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+
+    # Returns
+        A tensor.
+    """
     if data_format == 'channels_first':
         x = tf.transpose(x, (0, 4, 1, 2, 3))
 
@@ -2827,7 +3072,7 @@ def conv1d(x, kernel, strides=1, padding='valid',
         kernel: kernel tensor.
         strides: stride integer.
         padding: string, `"same"`, `"causal"` or `"valid"`.
-        data_format: string, one of "channels_last", "channels_first".
+        data_format: string, `"channels_last"` or `"channels_first"`.
         dilation_rate: integer dilate rate.
 
     # Returns
@@ -2863,9 +3108,9 @@ def conv2d(x, kernel, strides=(1, 1), padding='valid',
         kernel: kernel tensor.
         strides: strides tuple.
         padding: string, `"same"` or `"valid"`.
-        data_format: `"channels_last"` or `"channels_first"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
             Whether to use Theano or TensorFlow data format
-            for inputs/kernels/ouputs.
+            for inputs/kernels/outputs.
         dilation_rate: tuple of 2 integers.
 
     # Returns
@@ -2904,9 +3149,9 @@ def conv2d_transpose(x, kernel, output_shape, strides=(1, 1),
         output_shape: 1D int tensor for the output shape.
         strides: strides tuple.
         padding: string, `"same"` or `"valid"`.
-        data_format: `"channels_last"` or `"channels_first"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
             Whether to use Theano or TensorFlow data format
-            for inputs/kernels/ouputs.
+            for inputs/kernels/outputs.
 
     # Returns
         A tensor, result of transposed 2D convolution.
@@ -2941,8 +3186,8 @@ def separable_conv2d(x, depthwise_kernel, pointwise_kernel, strides=(1, 1),
         depthwise_kernel: convolution kernel for the depthwise convolution.
         pointwise_kernel: kernel for the 1x1 convolution.
         strides: strides tuple (length 2).
-        padding: padding mode, "valid" or "same".
-        data_format: data format, "channels_first" or "channels_last".
+        padding: string, `"same"` or `"valid"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
         dilation_rate: tuple of integers,
             dilation rates for the separable convolution.
 
@@ -2977,9 +3222,9 @@ def conv3d(x, kernel, strides=(1, 1, 1), padding='valid',
         kernel: kernel tensor.
         strides: strides tuple.
         padding: string, `"same"` or `"valid"`.
-        data_format: `"channels_last"` or `"channels_first"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
             Whether to use Theano or TensorFlow data format
-            for inputs/kernels/ouputs.
+            for inputs/kernels/outputs.
         dilation_rate: tuple of 3 integers.
 
     # Returns
@@ -3017,16 +3262,16 @@ def pool2d(x, pool_size, strides=(1, 1),
         x: Tensor or variable.
         pool_size: tuple of 2 integers.
         strides: tuple of 2 integers.
-        padding: one of `"valid"`, `"same"`.
-        data_format: one of `"channels_first"`, `"channels_last"`.
-        pool_mode: one of `"max"`, `"avg"`.
+        padding: string, `"same"` or `"valid"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+        pool_mode: string, `"max"` or `"avg"`.
 
     # Returns
         A tensor, result of 2D pooling.
 
     # Raises
-        ValueError: if `data_format` is neither `channels_last` or `channels_first`.
-        ValueError: if `pool_mode` is neither `max` or `avg`.
+        ValueError: if `data_format` is neither `"channels_last"` or `"channels_first"`.
+        ValueError: if `pool_mode` is neither `"max"` or `"avg"`.
     """
     if data_format is None:
         data_format = image_data_format()
@@ -3057,17 +3302,16 @@ def pool3d(x, pool_size, strides=(1, 1, 1), padding='valid',
         x: Tensor or variable.
         pool_size: tuple of 3 integers.
         strides: tuple of 3 integers.
-        padding: one of `"valid"`, `"same"`.
-        data_format: one of `"channels_first"`, `"channels_last"`.
-        pool_mode: one of `"max"`, `"avg"`.
+        padding: string, `"same"` or `"valid"`.
+        data_format: string, `"channels_last"` or `"channels_first"`.
+        pool_mode: string, `"max"` or `"avg"`.
 
     # Returns
         A tensor, result of 3D pooling.
 
     # Raises
-        ValueError: if `data_format` is neither
-            `channels_last` or `channels_first`.
-        ValueError: if `pool_mode` is neither `max` or `avg`.
+        ValueError: if `data_format` is neither `"channels_last"` or `"channels_first"`.
+        ValueError: if `pool_mode` is neither `"max"` or `"avg"`.
     """
     if data_format is None:
         data_format = image_data_format()
@@ -3096,35 +3340,60 @@ def bias_add(x, bias, data_format=None):
     # Arguments
         x: Tensor or variable.
         bias: Bias tensor to add.
-        data_format: Data format for 3D, 4D or 5D tensors:
-            one of "channels_first", "channels_last".
+        data_format: string, `"channels_last"` or `"channels_first"`.
 
     # Returns
         Output tensor.
 
     # Raises
-        ValueError: In case of invalid `data_format` argument.
+        ValueError: In one of the two cases below:
+                    1. invalid `data_format` argument.
+                    2. invalid bias shape.
+                       the bias should be either a vector or
+                       a tensor with ndim(x) - 1 dimension
     """
     if data_format is None:
         data_format = image_data_format()
     if data_format not in {'channels_first', 'channels_last'}:
         raise ValueError('Unknown data_format ' + str(data_format))
+    bias_shape = int_shape(bias)
+    if len(bias_shape) != 1 and len(bias_shape) != ndim(x) - 1:
+        raise ValueError('Unexpected bias dimensions %d, expect to be 1 or %d dimensions'
+                         % (len(bias_shape), ndim(x)))
     if ndim(x) == 5:
         if data_format == 'channels_first':
-            x += reshape(bias, (1, int_shape(bias)[0], 1, 1, 1))
+            if len(bias_shape) == 1:
+                x += reshape(bias, (1, bias_shape[0], 1, 1, 1))
+            else:
+                x += reshape(bias, (1, bias_shape[3]) + bias_shape[:3])
         elif data_format == 'channels_last':
-            x += reshape(bias, (1, 1, 1, 1, int_shape(bias)[0]))
+            if len(bias_shape) == 1:
+                x += reshape(bias, (1, 1, 1, bias_shape[0]))
+            else:
+                x += reshape(bias, (1,) + bias_shape)
     elif ndim(x) == 4:
         if data_format == 'channels_first':
-            x += reshape(bias, (1, int_shape(bias)[0], 1, 1))
+            if len(bias_shape) == 1:
+                x += reshape(bias, (1, bias_shape[0], 1, 1))
+            else:
+                x += reshape(bias, (1, bias_shape[2]) + bias_shape[:2])
         elif data_format == 'channels_last':
-            x = tf.nn.bias_add(x, bias,
-                               data_format='NHWC')
+            if len(bias_shape) == 1:
+                x = tf.nn.bias_add(x, bias,
+                                   data_format='NHWC')
+            else:
+                x += reshape(bias, (1,) + bias_shape)
     elif ndim(x) == 3:
         if data_format == 'channels_first':
-            x += reshape(bias, (1, int_shape(bias)[0], 1))
+            if len(bias_shape) == 1:
+                x += reshape(bias, (1, bias_shape[0], 1))
+            else:
+                x += reshape(bias, (1, bias_shape[1], bias_shape[0]))
         elif data_format == 'channels_last':
-            x += reshape(bias, (1, 1, int_shape(bias)[0]))
+            if len(bias_shape) == 1:
+                x += reshape(bias, (1, 1, bias_shape[0]))
+            else:
+                x += reshape(bias, (1, ) + bias_shape)
     else:
         x = tf.nn.bias_add(x, bias)
     return x
@@ -3389,3 +3658,110 @@ def foldr(fn, elems, initializer=None, name=None):
         Same type and shape as initializer
     """
     return tf.foldr(fn, elems, initializer=initializer, name=name)
+
+
+def local_conv1d(inputs, kernel, kernel_size, strides, data_format=None):
+    """Apply 1D conv with un-shared weights.
+
+    # Arguments
+        inputs: 3D tensor with shape: (batch_size, steps, input_dim)
+        kernel: the unshared weight for convolution,
+                with shape (output_length, feature_dim, filters)
+        kernel_size: a tuple of a single integer,
+                     specifying the length of the 1D convolution window
+        strides: a tuple of a single integer,
+                 specifying the stride length of the convolution
+        data_format: the data format, channels_first or channels_last
+
+    # Returns
+        the tensor after 1d conv with un-shared weights, with shape (batch_size, output_lenght, filters)
+
+    # Raises
+        ValueError: if `data_format` is neither `channels_last` or `channels_first`.
+    """
+    if data_format is None:
+        data_format = image_data_format()
+    if data_format not in {'channels_first', 'channels_last'}:
+        raise ValueError('Unknown data_format ' + str(data_format))
+
+    stride = strides[0]
+    kernel_shape = int_shape(kernel)
+    output_length, feature_dim, filters = kernel_shape
+
+    xs = []
+    for i in range(output_length):
+        slice_length = slice(i * stride,
+                             i * stride + kernel_size[0])
+        xs.append(reshape(inputs[:, slice_length, :],
+                          (1, -1, feature_dim)))
+    x_aggregate = concatenate(xs, axis=0)
+    # Shape: `(output_length, batch_size, filters)`.
+    output = batch_dot(x_aggregate, kernel)
+    return permute_dimensions(output, (1, 0, 2))
+
+
+def local_conv2d(inputs, kernel, kernel_size, strides, output_shape, data_format=None):
+    """Apply 2D conv with un-shared weights.
+
+    # Arguments
+        inputs: 4D tensor with shape:
+                (batch_size, filters, new_rows, new_cols)
+                if data_format='channels_first'
+                or 4D tensor with shape:
+                (batch_size, new_rows, new_cols, filters)
+                if data_format='channels_last'.
+        kernel: the unshared weight for convolution,
+                with shape (output_items, feature_dim, filters)
+        kernel_size: a tuple of 2 integers, specifying the
+                     width and height of the 2D convolution window.
+        strides: a tuple of 2 integers, specifying the strides
+                 of the convolution along the width and height.
+        output_shape: a tuple with (output_row, output_col)
+        data_format: the data format, channels_first or channels_last
+
+    # Returns
+        A 4d tensor with shape:
+        (batch_size, filters, new_rows, new_cols)
+        if data_format='channels_first'
+        or 4D tensor with shape:
+        (batch_size, new_rows, new_cols, filters)
+        if data_format='channels_last'.
+
+    # Raises
+        ValueError: if `data_format` is neither
+                    `channels_last` or `channels_first`.
+    """
+    if data_format is None:
+        data_format = image_data_format()
+    if data_format not in {'channels_first', 'channels_last'}:
+        raise ValueError('Unknown data_format ' + str(data_format))
+
+    stride_row, stride_col = strides
+    output_row, output_col = output_shape
+    kernel_shape = int_shape(kernel)
+    _, feature_dim, filters = kernel_shape
+
+    xs = []
+    for i in range(output_row):
+        for j in range(output_col):
+            slice_row = slice(i * stride_row,
+                              i * stride_row + kernel_size[0])
+            slice_col = slice(j * stride_col,
+                              j * stride_col + kernel_size[1])
+            if data_format == 'channels_first':
+                xs.append(reshape(inputs[:, :, slice_row, slice_col],
+                                  (1, -1, feature_dim)))
+            else:
+                xs.append(reshape(inputs[:, slice_row, slice_col, :],
+                                  (1, -1, feature_dim)))
+
+    x_aggregate = concatenate(xs, axis=0)
+    output = batch_dot(x_aggregate, kernel)
+    output = reshape(output,
+                     (output_row, output_col, -1, filters))
+
+    if data_format == 'channels_first':
+        output = permute_dimensions(output, (2, 3, 0, 1))
+    else:
+        output = permute_dimensions(output, (2, 0, 1, 3))
+    return output

keras/backend/theano_backend.py

@@ -1,6 +1,7 @@
 from collections import defaultdict
 from contextlib import contextmanager
 import theano
+from theano import ifelse
 from theano import tensor as T
 from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
 from theano.tensor.signal import pool
@@ -163,6 +164,42 @@ def constant(value, dtype=None, shape=None, name=None):
     return const
 
 
+def is_keras_tensor(x):
+    """Returns whether `x` is a Keras tensor.
+
+    # Arguments
+        x: a potential tensor.
+
+    # Returns
+        A boolean: whether the argument is a Keras tensor.
+
+    # Raises
+        ValueError: in case `x` is not a symbolic tensor.
+
+    # Examples
+    ```python
+        >>> from keras import backend as K
+        >>> np_var = numpy.array([1, 2])
+        >>> K.is_keras_tensor(np_var) # A numpy array is not a symbolic tensor.
+        ValueError
+        >>> k_var = theano.shared(value=np.array([1,2,3]))
+        >>> K.is_keras_tensor(k_var) # A variable created directly from tensorflow/theano is not a Keras tensor.
+        False
+        >>> keras_var = K.variable(np_var)
+        >>> K.is_keras_tensor(keras_var) # A variable created with the keras backend is a Keras tensor.
+        True
+        >>> keras_placeholder = K.placeholder(shape=(2, 4, 5))
+        >>> K.is_keras_tensor(keras_placeholder)  # A placeholder is a Keras tensor.
+        True
+    ```
+    """
+    if not isinstance(x, (T.TensorVariable,
+                          T.sharedvar.TensorSharedVariable)):
+        raise ValueError('Unexpectedly found an instance of type `' + str(type(x)) + '`. '
+                         'Expected a symbolic tensor instance.')
+    return hasattr(x, '_keras_history')
+
+
 def placeholder(shape=None, ndim=None, dtype=None, sparse=False, name=None):
     """Instantiate an input data placeholder variable.
     """
@@ -258,6 +295,18 @@ def zeros_like(x, dtype=None, name=None):
     return T.zeros_like(x, dtype=dtype)
 
 
+def identity(x):
+    """Returns a tensor with the same content as the input tensor.
+
+    # Arguments
+        x: The input tensor.
+
+    # Returns
+        A tensor of the same shape, type and content.
+    """
+    return x.copy()
+
+
 def random_uniform_variable(shape, low, high, dtype=None, name=None):
     return variable(np.random.uniform(low=low, high=high, size=shape),
                     dtype=dtype, name=name)
@@ -515,6 +564,29 @@ def log(x):
     return T.log(x)
 
 
+def logsumexp(x, axis=None, keepdims=False):
+    """Computes log(sum(exp(elements across dimensions of a tensor))).
+
+    This function is more numerically stable than log(sum(exp(x))).
+    It avoids overflows caused by taking the exp of large inputs and
+    underflows caused by taking the log of small inputs.
+
+    # Arguments
+        x: A tensor or variable.
+        axis: An integer, the axis to reduce over.
+        keepdims: A boolean, whether to keep the dimensions or not.
+            If `keepdims` is `False`, the rank of the tensor is reduced
+            by 1. If `keepdims` is `True`, the reduced dimension is
+            retained with length 1.
+
+    # Returns
+        The reduced tensor.
+    """
+    # Theano has a built-in optimization for logsumexp (see https://github.com/Theano/Theano/pull/4736)
+    # so we can just write the expression directly:
+    return T.log(T.sum(T.exp(x), axis=axis, keepdims=keepdims))
+
+
 def round(x):
     return T.round(x, mode='half_to_even')
 
@@ -1106,7 +1178,7 @@ def print_tensor(x, message=''):
 
 class Function(object):
 
-    def __init__(self, inputs, outputs, updates=[], **kwargs):
+    def __init__(self, inputs, outputs, updates=[], name=None, **kwargs):
         unique_variables_to_update = {}
         for v, nv in updates:
             if v not in unique_variables_to_update:
@@ -1115,7 +1187,9 @@ class Function(object):
         self.function = theano.function(inputs, outputs, updates=updates,
                                         allow_input_downcast=True,
                                         on_unused_input='ignore',
+                                        name=name,
                                         **kwargs)
+        self.name = name
 
     def __call__(self, inputs):
         assert isinstance(inputs, (list, tuple))
@@ -1127,7 +1201,7 @@ def function(inputs, outputs, updates=[], **kwargs):
         function_args = inspect.getargspec(theano.function)[0]
         for key in kwargs.keys():
             if key not in function_args:
-                msg = 'Invalid argument "%s" passed to K.function' % key
+                msg = 'Invalid argument "%s" passed to K.function with Theano backend' % key
                 raise ValueError(msg)
     return Function(inputs, outputs, updates=updates, **kwargs)
 
@@ -1488,8 +1562,9 @@ def dropout(x, level, noise_shape=None, seed=None):
     return x
 
 
-def l2_normalize(x, axis):
-    norm = T.sqrt(T.sum(T.square(x), axis=axis, keepdims=True))
+def l2_normalize(x, axis, epsilon=1e-12):
+    square_sum = T.sum(T.square(x), axis=axis, keepdims=True)
+    norm = T.sqrt(T.maximum(square_sum, epsilon))
     return x / norm
 
 
@@ -1739,7 +1814,7 @@ def conv2d(x, kernel, strides=(1, 1), padding='valid',
         padding: string, "same" or "valid".
         data_format: "channels_last" or "channels_first".
             Whether to use Theano or TensorFlow data format
-        in inputs/kernels/ouputs.
+        in inputs/kernels/outputs.
     """
     if data_format is None:
         data_format = image_data_format()
@@ -1783,7 +1858,10 @@ def conv2d_transpose(x, kernel, output_shape, strides=(1, 1),
         padding: string, "same" or "valid".
         data_format: "channels_last" or "channels_first".
             Whether to use Theano or TensorFlow data format
-        in inputs/kernels/ouputs.
+        in inputs/kernels/outputs.
+
+    # Raises
+        ValueError: if using an even kernel size with padding 'same'.
     """
     flip_filters = False
     if data_format is None:
@@ -1802,6 +1880,12 @@ def conv2d_transpose(x, kernel, output_shape, strides=(1, 1),
     else:
         # Will only work if `kernel` is a shared variable.
         kernel_shape = kernel.eval().shape
+
+    if padding == 'same' and kernel_shape[0] % 2 == 0:
+        raise ValueError('In `Conv2DTranspose`, with padding mode `same`, '
+                         'even kernel sizes are only supported with Tensorflow. '
+                         'With Theano, set `kernel_size` to an odd number.')
+
     kernel_shape = _preprocess_conv2d_filter_shape(kernel_shape, data_format)
 
     x = _preprocess_conv2d_input(x, data_format)
@@ -1835,7 +1919,7 @@ def conv3d(x, kernel, strides=(1, 1, 1),
         padding: string, "same" or "valid".
         data_format: "channels_last" or "channels_first".
             Whether to use Theano or TensorFlow data format
-        in inputs/kernels/ouputs.
+        in inputs/kernels/outputs.
     """
     if data_format is None:
         data_format = image_data_format()
@@ -1898,10 +1982,14 @@ def pool2d(x, pool_size, strides=(1, 1), padding='valid',
                                 pad=pad,
                                 mode='max')
     elif pool_mode == 'avg':
+        if padding == 'same':
+            th_avg_pool_mode = 'average_inc_pad'
+        elif padding == 'valid':
+            th_avg_pool_mode = 'average_exc_pad'
         pool_out = pool.pool_2d(x, ws=pool_size, stride=strides,
                                 ignore_border=True,
                                 pad=pad,
-                                mode='average_exc_pad')
+                                mode=th_avg_pool_mode)
     else:
         raise ValueError('Invalid pooling mode:', pool_mode)
     if padding == 'same':
@@ -1971,21 +2059,44 @@ def bias_add(x, bias, data_format=None):
         data_format = image_data_format()
     if data_format not in {'channels_first', 'channels_last'}:
         raise ValueError('Unknown data_format ' + str(data_format))
+    if ndim(bias) != 1 and ndim(bias) != ndim(x) - 1:
+        raise ValueError('Unexpected bias dimensions %d, '
+                         'expect to be 1 or %d dimensions'
+                         % (ndim(bias), ndim(x) - 1))
+    bias_shape = tuple(bias.shape)
     if ndim(x) == 5:
         if data_format == 'channels_first':
-            x += reshape(bias, (1, bias.shape[0], 1, 1, 1))
+            if ndim(bias) == 1:
+                x += reshape(bias, (1, bias_shape[0], 1, 1, 1))
+            else:
+                x += reshape(bias, (1, bias_shape[3]) + bias_shape[:3])
         elif data_format == 'channels_last':
-            x += reshape(bias, (1, 1, 1, 1, bias.shape[0]))
+            if ndim(bias) == 1:
+                x += reshape(bias, (1, 1, 1, 1, bias_shape[0]))
+            else:
+                x += reshape(bias, (1,) + bias_shape)
     elif ndim(x) == 4:
         if data_format == 'channels_first':
-            x += reshape(bias, (1, bias.shape[0], 1, 1))
+            if ndim(bias) == 1:
+                x += reshape(bias, (1, bias_shape[0], 1, 1))
+            else:
+                x += reshape(bias, (1, bias_shape[2]) + bias_shape[:2])
         elif data_format == 'channels_last':
-            x += reshape(bias, (1, 1, 1, bias.shape[0]))
+            if ndim(bias) == 1:
+                x += reshape(bias, (1, 1, 1, bias_shape[0]))
+            else:
+                x += reshape(bias, (1,) + bias_shape)
     elif ndim(x) == 3:
         if data_format == 'channels_first':
-            x += reshape(bias, (1, bias.shape[0], 1))
+            if ndim(bias) == 1:
+                x += reshape(bias, (1, bias_shape[0], 1))
+            else:
+                x += reshape(bias, (1, bias_shape[1], bias_shape[0]))
         elif data_format == 'channels_last':
-            x += reshape(bias, (1, 1, bias.shape[0]))
+            if ndim(bias) == 1:
+                x += reshape(bias, (1, 1, bias_shape[0]))
+            else:
+                x += reshape(bias, (1,) + bias_shape)
     else:
         x += bias
     return x
@@ -2203,3 +2314,72 @@ def foldr(fn, elems, initializer=None, name=None):
     fn2 = lambda x, acc: fn(acc, x)
 
     return theano.foldr(fn2, elems, initializer, name=name)[0]
+
+
+def local_conv1d(inputs, kernel, kernel_size, strides, data_format=None):
+    if data_format is None:
+        data_format = image_data_format()
+    if data_format not in {'channels_first', 'channels_last'}:
+        raise ValueError('Unknown data_format ' + str(data_format))
+
+    stride = strides[0]
+    kernel_shape = int_shape(kernel)
+    output_length, feature_dim, filters = kernel_shape
+
+    xs = []
+    for i in range(output_length):
+        slice_length = slice(i * stride,
+                             i * stride + kernel_size[0])
+        xs.append(reshape(inputs[:, slice_length, :],
+                          (1, -1, feature_dim)))
+    x_aggregate = concatenate(xs, axis=0)
+    # Shape: `(output_length, batch_size, filters)`.
+    output = batch_dot(x_aggregate, kernel)
+    return permute_dimensions(output, (1, 0, 2))
+
+
+def local_conv2d(inputs, kernel, kernel_size, strides, output_shape, data_format=None):
+    if data_format is None:
+        data_format = image_data_format()
+    if data_format not in {'channels_first', 'channels_last'}:
+        raise ValueError('Unknown data_format ' + str(data_format))
+
+    stride_row, stride_col = strides
+    output_row, output_col = output_shape
+    kernel_shape = int_shape(kernel)
+    _, feature_dim, filters = kernel_shape
+
+    if data_format == 'channels_first':
+        output = []
+        for i in range(output_row):
+            for j in range(output_col):
+                slice_row = slice(i * stride_row,
+                                  i * stride_row + kernel_size[0])
+                slice_col = slice(j * stride_col,
+                                  j * stride_col + kernel_size[1])
+                x_flatten = reshape(inputs[:, :, slice_row, slice_col],
+                                    (1, -1, feature_dim))
+                output.append(dot(x_flatten,
+                                  kernel[i * output_col + j, :, :]))
+        output = concatenate(output, axis=0)
+        output = reshape(output,
+                         (output_row, output_col, -1, filters))
+        output = permute_dimensions(output, (2, 3, 0, 1))
+    else:
+        xs = []
+        for i in range(output_row):
+            for j in range(output_col):
+                slice_row = slice(i * stride_row,
+                                  i * stride_row + kernel_size[0])
+                slice_col = slice(j * stride_col,
+                                  j * stride_col + kernel_size[1])
+                xs.append(reshape(inputs[:, slice_row, slice_col, :],
+                                  (1, -1, feature_dim)))
+
+        x_aggregate = concatenate(xs, axis=0)
+        output = batch_dot(x_aggregate, kernel)
+        output = reshape(output,
+                         (output_row, output_col, -1, filters))
+        output = permute_dimensions(output, (2, 0, 1, 3))
+
+    return output

keras/callbacks.py

@@ -3,6 +3,7 @@ from __future__ import print_function
 
 import os
 import csv
+import six
 
 import numpy as np
 import time
@@ -226,6 +227,21 @@ class BaseLogger(Callback):
                     logs[k] = self.totals[k] / self.seen
 
 
+class TerminateOnNaN(Callback):
+    """Callback that terminates training when a NaN loss is encountered."""
+
+    def __init__(self):
+        super(TerminateOnNaN, self).__init__()
+
+    def on_batch_end(self, batch, logs=None):
+        logs = logs or {}
+        loss = logs.get('loss')
+        if loss is not None:
+            if np.isnan(loss) or np.isinf(loss):
+                print('Batch %d: Invalid loss, terminating training' % (batch))
+                self.model.stop_training = True
+
+
 class ProgbarLogger(Callback):
     """Callback that prints metrics to stdout.
 
@@ -466,7 +482,9 @@ class EarlyStopping(Callback):
             self.min_delta *= -1
 
     def on_train_begin(self, logs=None):
-        self.wait = 0  # Allow instances to be re-used
+        # Allow instances to be re-used
+        self.wait = 0
+        self.stopped_epoch = 0
         self.best = np.Inf if self.monitor_op == np.less else -np.Inf
 
     def on_epoch_end(self, epoch, logs=None):
@@ -503,8 +521,7 @@ class RemoteMonitor(Callback):
         field: String; JSON field under which the data will be stored.
         headers: Dictionary; optional custom HTTP headers.
             Defaults to:
-            `{'Accept': 'application/json',
-              'Content-Type': 'application/json'}`
+            `{'Accept': 'application/json', 'Content-Type': 'application/json'}`
     """
 
     def __init__(self,
@@ -578,19 +595,24 @@ class TensorBoard(Callback):
     tensorboard --logdir=/full_path_to_your_logs
     ```
     You can find more information about TensorBoard
-    [here](https://www.tensorflow.org/versions/master/how_tos/summaries_and_tensorboard/index.html).
+    [here](https://www.tensorflow.org/get_started/summaries_and_tensorboard).
 
     # Arguments
         log_dir: the path of the directory where to save the log
-            files to be parsed by Tensorboard.
+            files to be parsed by TensorBoard.
         histogram_freq: frequency (in epochs) at which to compute activation
-            histograms for the layers of the model. If set to 0,
-            histograms won't be computed.
-        write_graph: whether to visualize the graph in Tensorboard.
+            and weight histograms for the layers of the model. If set to 0,
+            histograms won't be computed. Validation data (or split) must be
+            specified for histogram visualizations.
+        write_graph: whether to visualize the graph in TensorBoard.
             The log file can become quite large when
             write_graph is set to True.
+        write_grads: whether to visualize gradient histograms in TensorBoard.
+            `histogram_freq` must be greater than 0.
+        batch_size: size of batch of inputs to feed to the network
+            for histograms computation.
         write_images: whether to write model weights to visualize as
-            image in Tensorboard.
+            image in TensorBoard.
         embeddings_freq: frequency (in epochs) at which selected embedding
             layers will be saved.
         embeddings_layer_names: a list of names of layers to keep eye on. If
@@ -604,7 +626,9 @@ class TensorBoard(Callback):
 
     def __init__(self, log_dir='./logs',
                  histogram_freq=0,
+                 batch_size=32,
                  write_graph=True,
+                 write_grads=False,
                  write_images=False,
                  embeddings_freq=0,
                  embeddings_layer_names=None,
@@ -617,10 +641,12 @@ class TensorBoard(Callback):
         self.histogram_freq = histogram_freq
         self.merged = None
         self.write_graph = write_graph
+        self.write_grads = write_grads
         self.write_images = write_images
         self.embeddings_freq = embeddings_freq
         self.embeddings_layer_names = embeddings_layer_names
         self.embeddings_metadata = embeddings_metadata or {}
+        self.batch_size = batch_size
 
     def set_model(self, model):
         self.model = model
@@ -630,14 +656,42 @@ class TensorBoard(Callback):
 
                 for weight in layer.weights:
                     tf.summary.histogram(weight.name, weight)
+                    if self.write_grads:
+                        grads = model.optimizer.get_gradients(model.total_loss,
+                                                              weight)
+                        tf.summary.histogram('{}_grad'.format(weight.name), grads)
                     if self.write_images:
                         w_img = tf.squeeze(weight)
-                        shape = w_img.get_shape()
-                        if len(shape) > 1 and shape[0] > shape[1]:
-                            w_img = tf.transpose(w_img)
-                        if len(shape) == 1:
-                            w_img = tf.expand_dims(w_img, 0)
-                        w_img = tf.expand_dims(tf.expand_dims(w_img, 0), -1)
+                        shape = K.int_shape(w_img)
+                        if len(shape) == 2:  # dense layer kernel case
+                            if shape[0] > shape[1]:
+                                w_img = tf.transpose(w_img)
+                                shape = K.int_shape(w_img)
+                            w_img = tf.reshape(w_img, [1,
+                                                       shape[0],
+                                                       shape[1],
+                                                       1])
+                        elif len(shape) == 3:  # convnet case
+                            if K.image_data_format() == 'channels_last':
+                                # switch to channels_first to display
+                                # every kernel as a separate image
+                                w_img = tf.transpose(w_img, perm=[2, 0, 1])
+                                shape = K.int_shape(w_img)
+                            w_img = tf.reshape(w_img, [shape[0],
+                                                       shape[1],
+                                                       shape[2],
+                                                       1])
+                        elif len(shape) == 1:  # bias case
+                            w_img = tf.reshape(w_img, [1,
+                                                       shape[0],
+                                                       1,
+                                                       1])
+                        else:
+                            # not possible to handle 3D convnets etc.
+                            continue
+
+                        shape = K.int_shape(w_img)
+                        assert len(shape) == 4 and shape[-1] in [1, 3, 4]
                         tf.summary.image(weight.name, w_img)
 
                 if hasattr(layer, 'output'):
@@ -652,8 +706,6 @@ class TensorBoard(Callback):
             self.writer = tf.summary.FileWriter(self.log_dir)
 
         if self.embeddings_freq:
-            self.saver = tf.train.Saver()
-
             embeddings_layer_names = self.embeddings_layer_names
 
             if not embeddings_layer_names:
@@ -664,6 +716,8 @@ class TensorBoard(Callback):
                           for layer in self.model.layers
                           if layer.name in embeddings_layer_names}
 
+            self.saver = tf.train.Saver(list(embeddings.values()))
+
             embeddings_metadata = {}
 
             if not isinstance(self.embeddings_metadata, str):
@@ -673,15 +727,13 @@ class TensorBoard(Callback):
                                        for layer_name in embeddings.keys()}
 
             config = projector.ProjectorConfig()
-            self.embeddings_logs = []
+            self.embeddings_ckpt_path = os.path.join(self.log_dir,
+                                                     'keras_embedding.ckpt')
 
             for layer_name, tensor in embeddings.items():
                 embedding = config.embeddings.add()
                 embedding.tensor_name = tensor.name
 
-                self.embeddings_logs.append(os.path.join(self.log_dir,
-                                                         layer_name + '.ckpt'))
-
                 if layer_name in embeddings_metadata:
                     embedding.metadata_path = embeddings_metadata[layer_name]
 
@@ -692,24 +744,37 @@ class TensorBoard(Callback):
 
         if self.validation_data and self.histogram_freq:
             if epoch % self.histogram_freq == 0:
-                # TODO: implement batched calls to sess.run
-                # (current call will likely go OOM on GPU)
-                if self.model.uses_learning_phase:
-                    cut_v_data = len(self.model.inputs)
-                    val_data = self.validation_data[:cut_v_data] + [0]
-                    tensors = self.model.inputs + [K.learning_phase()]
-                else:
-                    val_data = self.validation_data
-                    tensors = self.model.inputs
-                feed_dict = dict(zip(tensors, val_data))
-                result = self.sess.run([self.merged], feed_dict=feed_dict)
-                summary_str = result[0]
-                self.writer.add_summary(summary_str, epoch)
 
-        if self.embeddings_freq and self.embeddings_logs:
+                val_data = self.validation_data
+                tensors = (self.model.inputs +
+                           self.model.targets +
+                           self.model.sample_weights)
+
+                if self.model.uses_learning_phase:
+                    tensors += [K.learning_phase()]
+
+                assert len(val_data) == len(tensors)
+                val_size = val_data[0].shape[0]
+                i = 0
+                while i < val_size:
+                    step = min(self.batch_size, val_size - i)
+                    batch_val = []
+                    batch_val.append(val_data[0][i:i + step])
+                    batch_val.append(val_data[1][i:i + step])
+                    batch_val.append(val_data[2][i:i + step])
+                    if self.model.uses_learning_phase:
+                        batch_val.append(val_data[3])
+                    feed_dict = dict(zip(tensors, batch_val))
+                    result = self.sess.run([self.merged], feed_dict=feed_dict)
+                    summary_str = result[0]
+                    self.writer.add_summary(summary_str, epoch)
+                    i += self.batch_size
+
+        if self.embeddings_freq and self.embeddings_ckpt_path:
             if epoch % self.embeddings_freq == 0:
-                for log in self.embeddings_logs:
-                    self.saver.save(self.sess, log, epoch)
+                self.saver.save(self.sess,
+                                self.embeddings_ckpt_path,
+                                epoch)
 
         for name, value in logs.items():
             if name in ['batch', 'size']:
@@ -735,9 +800,9 @@ class ReduceLROnPlateau(Callback):
 
     # Example
         ```python
-            reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2,
-                                          patience=5, min_lr=0.001)
-            model.fit(X_train, Y_train, callbacks=[reduce_lr])
+        reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2,
+                                      patience=5, min_lr=0.001)
+        model.fit(X_train, Y_train, callbacks=[reduce_lr])
         ```
 
     # Arguments
@@ -844,8 +909,8 @@ class CSVLogger(Callback):
 
     # Example
         ```python
-            csv_logger = CSVLogger('training.log')
-            model.fit(X_train, Y_train, callbacks=[csv_logger])
+        csv_logger = CSVLogger('training.log')
+        model.fit(X_train, Y_train, callbacks=[csv_logger])
         ```
 
     # Arguments
@@ -862,23 +927,26 @@ class CSVLogger(Callback):
         self.writer = None
         self.keys = None
         self.append_header = True
+        self.file_flags = 'b' if six.PY2 and os.name == 'nt' else ''
         super(CSVLogger, self).__init__()
 
     def on_train_begin(self, logs=None):
         if self.append:
             if os.path.exists(self.filename):
-                with open(self.filename) as f:
+                with open(self.filename, 'r' + self.file_flags) as f:
                     self.append_header = not bool(len(f.readline()))
-            self.csv_file = open(self.filename, 'a')
+            self.csv_file = open(self.filename, 'a' + self.file_flags)
         else:
-            self.csv_file = open(self.filename, 'w')
+            self.csv_file = open(self.filename, 'w' + self.file_flags)
 
     def on_epoch_end(self, epoch, logs=None):
         logs = logs or {}
 
         def handle_value(k):
             is_zero_dim_ndarray = isinstance(k, np.ndarray) and k.ndim == 0
-            if isinstance(k, Iterable) and not is_zero_dim_ndarray:
+            if isinstance(k, six.string_types):
+                return k
+            elif isinstance(k, Iterable) and not is_zero_dim_ndarray:
                 return '"[%s]"' % (', '.join(map(str, k)))
             else:
                 return k
@@ -910,6 +978,7 @@ class LambdaCallback(Callback):
     This callback is constructed with anonymous functions that will be called
     at the appropriate time. Note that the callbacks expects positional
     arguments, as:
+
      - `on_epoch_begin` and `on_epoch_end` expect two positional arguments:
         `epoch`, `logs`
      - `on_batch_begin` and `on_batch_end` expect two positional arguments:

keras/constraints.py

@@ -58,7 +58,7 @@ class NonNeg(Constraint):
     """
 
     def __call__(self, w):
-        w *= K.cast(w >= 0., K.floatx())
+        w *= K.cast(K.greater_equal(w, 0.), K.floatx())
         return w
 
 

keras/datasets/boston_housing.py

@@ -16,7 +16,9 @@ def load_data(path='boston_housing.npz', seed=113, test_split=0.2):
         Tuple of Numpy arrays: `(x_train, y_train), (x_test, y_test)`.
     """
     assert 0 <= test_split < 1
-    path = get_file(path, origin='https://s3.amazonaws.com/keras-datasets/boston_housing.npz')
+    path = get_file(path,
+                    origin='https://s3.amazonaws.com/keras-datasets/boston_housing.npz',
+                    file_hash='f553886a1f8d56431e820c5b82552d9d95cfcb96d1e678153f8839538947dff5')
     f = np.load(path)
     x = f['x']
     y = f['y']

keras/datasets/imdb.py

@@ -100,7 +100,7 @@ def load_data(path='imdb.npz', num_words=None, skip_top=0,
         for x in xs:
             nx = []
             for w in x:
-                if w >= num_words or w < skip_top:
+                if skip_top <= w < num_words:
                     nx.append(w)
             new_xs.append(nx)
         xs = new_xs

keras/datasets/reuters.py

@@ -84,7 +84,7 @@ def load_data(path='reuters.npz', num_words=None, skip_top=0,
         for x in xs:
             nx = []
             for w in x:
-                if w >= num_words or w < skip_top:
+                if skip_top <= w < num_words:
                     nx.append(w)
             new_xs.append(nx)
         xs = new_xs

keras/engine/topology.py

@@ -360,28 +360,35 @@ class Layer(object):
     def non_trainable_weights(self, weights):
         self._non_trainable_weights = weights
 
-    def add_weight(self, shape, initializer,
-                   name=None,
-                   trainable=True,
+    @interfaces.legacy_add_weight_support
+    def add_weight(self,
+                   name,
+                   shape,
+                   dtype=None,
+                   initializer=None,
                    regularizer=None,
+                   trainable=True,
                    constraint=None):
         """Adds a weight variable to the layer.
 
         # Arguments
-            shape: The shape tuple of the weight.
-            initializer: An Initializer instance (callable).
             name: String, the name for the weight variable.
+            shape: The shape tuple of the weight.
+            dtype: The dtype of the weight.
+            initializer: An Initializer instance (callable).
+            regularizer: An optional Regularizer instance.
             trainable: A boolean, whether the weight should
                 be trained via backprop or not (assuming
                 that the layer itself is also trainable).
-            regularizer: An optional Regularizer instance.
             constraint: An optional Constraint instance.
 
         # Returns
             The created weight variable.
         """
         initializer = initializers.get(initializer)
-        weight = K.variable(initializer(shape), dtype=K.floatx(), name=name)
+        if dtype is None:
+            dtype = K.floatx()
+        weight = K.variable(initializer(shape), dtype=dtype, name=name)
         if regularizer is not None:
             self.add_loss(regularizer(weight))
         if constraint is not None:
@@ -406,13 +413,24 @@ class Layer(object):
             ValueError: in case of mismatch between
                 the provided inputs and the expectations of the layer.
         """
+        inputs = _to_list(inputs)
+        for x in inputs:
+            try:
+                K.is_keras_tensor(x)
+            except ValueError:
+                raise ValueError('Layer ' + self.name + ' was called with '
+                                 'an input that isn\'t a symbolic tensor. '
+                                 'Received type: ' +
+                                 str(type(x)) + '. Full input: ' +
+                                 str(inputs) + '. All inputs to the layer '
+                                 'should be tensors.')
+
         if not self.input_spec:
             return
         if not isinstance(self.input_spec, (list, tuple)):
             input_spec = _to_list(self.input_spec)
         else:
             input_spec = self.input_spec
-        inputs = _to_list(inputs)
         if len(inputs) != len(input_spec):
             raise ValueError('Layer ' + self.name + ' expects ' +
                              str(len(input_spec)) + ' inputs, '
@@ -578,6 +596,20 @@ class Layer(object):
             output = self.call(inputs, **kwargs)
             output_mask = self.compute_mask(inputs, previous_mask)
 
+            # If the layer returns tensors from its inputs, unmodified,
+            # we copy them to avoid loss of tensor metadata.
+            output_ls = _to_list(output)
+            inputs_ls = _to_list(inputs)
+            output_ls_copy = []
+            for x in output_ls:
+                if x in inputs_ls:
+                    x = K.identity(x)
+                output_ls_copy.append(x)
+            if len(output_ls_copy) == 1:
+                output = output_ls_copy[0]
+            else:
+                output = output_ls_copy
+
             # Infering the output shape is only relevant for Theano.
             if all([s is not None for s in _to_list(input_shape)]):
                 output_shape = self.compute_output_shape(input_shape)
@@ -1056,7 +1088,7 @@ class Layer(object):
         if hasattr(self, '_losses'):
             self._losses += losses
         # Update self._per_input_updates
-        if inputs == []:
+        if isinstance(input, list) and inputs == []:
             inputs = None
         if inputs is not None:
             inputs_hash = _object_list_uid(inputs)
@@ -1088,7 +1120,7 @@ class Layer(object):
         if hasattr(self, '_updates'):
             self._updates += updates
         # Update self._per_input_updates
-        if inputs == []:
+        if isinstance(inputs, list) and inputs == []:
             inputs = None
         if inputs is not None:
             inputs_hash = _object_list_uid(inputs)
@@ -1249,6 +1281,7 @@ class InputLayer(Layer):
         name: Name of the layer (string).
     """
 
+    @interfaces.legacy_input_support
     def __init__(self, input_shape=None, batch_size=None,
                  batch_input_shape=None,
                  dtype=None, input_tensor=None, sparse=False, name=None):
@@ -1265,7 +1298,8 @@ class InputLayer(Layer):
             raise ValueError('Only provide the input_shape OR '
                              'batch_input_shape argument to '
                              'InputLayer, not both at the same time.')
-        if input_tensor is not None:
+        if input_tensor is not None and batch_input_shape is None:
+            # If input_tensor is set, and batch_input_shape is not set:
             # Attempt automatic input shape inference.
             try:
                 batch_input_shape = K.int_shape(input_tensor)
@@ -1337,7 +1371,7 @@ def Input(shape=None, batch_shape=None,
     attributes that allow us to build a Keras model
     just by knowing the inputs and outputs of the model.
 
-    For instance, if a, b and c and Keras tensors,
+    For instance, if a, b and c are Keras tensors,
     it becomes possible to do:
     `model = Model(input=[a, b], output=c)`
 
@@ -1434,6 +1468,9 @@ class Container(Layer):
 
     # Class Methods
         from_config
+
+    # Raises
+        TypeError: if input tensors are not Keras tensors from InputLayer objects
     """
 
     @interfaces.legacy_model_constructor_support
@@ -1460,13 +1497,19 @@ class Container(Layer):
             self.outputs = [outputs]
 
         # Check for redundancy in inputs.
-        inputs_set = set(self.inputs)
-        if len(inputs_set) != len(self.inputs):
+        if len(set(self.inputs)) != len(self.inputs):
             raise ValueError('The list of inputs passed to the model '
                              'is redundant. '
                              'All inputs should only appear once.'
                              ' Found: ' + str(self.inputs))
 
+        # Check for redundancy in outputs.
+        if len(set(self.outputs)) != len(self.outputs):
+            warnings.warn('The list of outputs passed to the model '
+                          'is redundant. '
+                          'All outputs should only appear once.'
+                          ' Found: ' + str(self.outputs))
+
         # List of initial layers (1 to 1 mapping with self.inputs,
         # hence the same layer might appear twice)
         self.input_layers = []
@@ -1568,6 +1611,15 @@ class Container(Layer):
         self._feed_inputs = []
         self._feed_input_shapes = []
         for i, layer in enumerate(self.input_layers):
+            # Check that layer is an InputLayer.
+            if not isinstance(layer, InputLayer):
+                raise TypeError(
+                    'Input layers to a `Model` must be `InputLayer` objects. '
+                    'Received inputs: {}. '
+                    'Input {} (0-based) originates '
+                    'from layer type `{}`.'.format(inputs,
+                                                   i,
+                                                   layer.__class__.__name__))
             self.input_names.append(layer.name)
             if layer.is_placeholder:
                 self._feed_input_names.append(layer.name)
@@ -1586,72 +1638,92 @@ class Container(Layer):
         nodes_depths = {}  # dict {node: depth value}
         layers_depths = {}  # dict {layer: depth value}
         layer_indices = {}  # dict {layer: index in traversal}
+        nodes_in_decreasing_depth = []
 
-        def make_node_marker(node, depth):
-            return str(id(node)) + '-' + str(depth)
-
-        def build_map_of_graph(tensor, seen_nodes=None, depth=0,
+        def build_map_of_graph(tensor, finished_nodes, nodes_in_progress,
                                layer=None, node_index=None, tensor_index=None):
             """Builds a map of the graph of layers.
 
-            This recursively updates the maps `nodes_depths`,
-            `layers_depths` and the set `container_nodes`.
-
-            Does not try to detect cycles in the graph.
+            This recursively updates the map `layer_indices`,
+            the list `nodes_in_decreasing_depth` and the set `container_nodes`.
 
             # Arguments
                 tensor: Some tensor in a graph.
-                seen_nodes: Set of node ids ("{layer.name}_ib-{node_index}")
-                    of nodes seen so far. Useful to prevent infinite loops.
-                depth: Current depth in the graph (0 = last output).
+                finished_nodes: Set of nodes whose subgraphs have been traversed
+                    completely. Useful to prevent duplicated work.
+                nodes_in_progress: Set of nodes that are currently active on the
+                    recursion stack. Useful to detect cycles.
                 layer: Layer from which `tensor` comes from. If not provided,
                     will be obtained from `tensor._keras_history`.
                 node_index: Node index from which `tensor` comes from.
                 tensor_index: Tensor_index from which `tensor` comes from.
+
+            # Raises
+                RuntimeError: if a cycle is detected.
             """
-            seen_nodes = seen_nodes or set()
             if not layer or node_index is None or tensor_index is None:
                 layer, node_index, tensor_index = tensor._keras_history
             node = layer.inbound_nodes[node_index]
 
             # Prevent cycles.
-            seen_nodes.add(make_node_marker(node, depth))
+            if node in nodes_in_progress:
+                raise RuntimeError(
+                    'The tensor ' + str(tensor) + ' at layer "' +
+                    layer.name + '" is part of a cycle.')
+
+            # Don't repeat work for shared subgraphs
+            if node in finished_nodes:
+                return
 
             node_key = layer.name + '_ib-' + str(node_index)
             # Update container_nodes.
             container_nodes.add(node_key)
-            # Update nodes_depths.
-            node_depth = nodes_depths.get(node)
-            if node_depth is None:
-                nodes_depths[node] = depth
-            else:
-                nodes_depths[node] = max(depth, node_depth)
-            # Update layers_depths.
-            previously_seen_depth = layers_depths.get(layer)
-            if previously_seen_depth is None:
-                current_depth = depth
-            else:
-                current_depth = max(depth, previously_seen_depth)
-            layers_depths[layer] = current_depth
+
+            # Store the traversal order for layer sorting.
             if layer not in layer_indices:
                 layer_indices[layer] = len(layer_indices)
 
+            nodes_in_progress.add(node)
+
             # Propagate to all previous tensors connected to this node.
             for i in range(len(node.inbound_layers)):
                 x = node.input_tensors[i]
                 layer = node.inbound_layers[i]
                 node_index = node.node_indices[i]
                 tensor_index = node.tensor_indices[i]
-                next_node = layer.inbound_nodes[node_index]
-                # use node_marker to prevent cycles
-                node_marker = make_node_marker(next_node, current_depth + 1)
-                if node_marker not in seen_nodes:
-                    build_map_of_graph(x, seen_nodes, current_depth + 1,
-                                       layer, node_index, tensor_index)
+                build_map_of_graph(x, finished_nodes, nodes_in_progress,
+                                   layer, node_index, tensor_index)
 
+            finished_nodes.add(node)
+            nodes_in_progress.remove(node)
+
+            nodes_in_decreasing_depth.append(node)
+
+        finished_nodes = set()
+        nodes_in_progress = set()
         for x in self.outputs:
-            seen_nodes = set()
-            build_map_of_graph(x, seen_nodes, depth=0)
+            build_map_of_graph(x, finished_nodes, nodes_in_progress)
+
+        for node in reversed(nodes_in_decreasing_depth):
+            # If the depth is not set, the node has no outbound nodes (depth 0).
+            depth = nodes_depths.setdefault(node, 0)
+
+            # Update the depth of the corresponding layer
+            previous_depth = layers_depths.get(node.outbound_layer, 0)
+            # If we've seen this layer before at a higher depth, we should use that depth instead
+            # of the node depth.  This is necessary for shared layers that have inputs at different
+            # depth levels in the graph.
+            depth = max(depth, previous_depth)
+            layers_depths[node.outbound_layer] = depth
+            nodes_depths[node] = depth
+
+            # Update the depth of inbound nodes.
+            for i in range(len(node.inbound_layers)):
+                inbound_layer = node.inbound_layers[i]
+                node_index = node.node_indices[i]
+                inbound_node = inbound_layer.inbound_nodes[node_index]
+                previous_depth = nodes_depths.get(inbound_node, 0)
+                nodes_depths[inbound_node] = max(depth + 1, previous_depth)
 
         # Build a dict {depth: list of nodes with this depth}
         nodes_by_depth = {}
@@ -2747,6 +2819,25 @@ def preprocess_weights_for_loading(layer, weights,
         A list of weights values (Numpy arrays).
     """
     if original_keras_version == '1':
+        if layer.__class__.__name__ == 'Bidirectional':
+            num_weights_per_layer = len(weights) // 2
+
+            forward_weights = preprocess_weights_for_loading(layer.forward_layer,
+                                                             weights[:num_weights_per_layer],
+                                                             original_keras_version,
+                                                             original_backend)
+            backward_weights = preprocess_weights_for_loading(layer.backward_layer,
+                                                              weights[num_weights_per_layer:],
+                                                              original_keras_version,
+                                                              original_backend)
+            weights = forward_weights + backward_weights
+
+        if layer.__class__.__name__ == 'TimeDistributed':
+            weights = preprocess_weights_for_loading(layer.layer,
+                                                     weights,
+                                                     original_keras_version,
+                                                     original_backend)
+
         if layer.__class__.__name__ == 'Conv1D':
             shape = weights[0].shape
             # Handle Keras 1.1 format
@@ -2832,16 +2923,20 @@ def preprocess_weights_for_loading(layer, weights,
                                                     (2, 3, 1, 0))
                 weights = [kernel, recurrent_kernel, bias]
 
-    if original_backend and K.backend() != original_backend:
-        conv_layers = ['Conv1D',
-                       'Conv2D',
-                       'Conv3D',
-                       'Conv2DTranspose']
-        if layer.__class__.__name__ in conv_layers:
+    conv_layers = ['Conv1D',
+                   'Conv2D',
+                   'Conv3D',
+                   'Conv2DTranspose',
+                   'ConvLSTM2D']
+    if layer.__class__.__name__ in conv_layers:
+        if original_backend and K.backend() != original_backend:
             weights[0] = conv_utils.convert_kernel(weights[0])
-        if layer.__class__.__name__ == 'ConvLSTM2D':
-            weights[0] = conv_utils.convert_kernel(weights[0])
-            weights[1] = conv_utils.convert_kernel(weights[1])
+            if layer.__class__.__name__ == 'ConvLSTM2D':
+                weights[1] = conv_utils.convert_kernel(weights[1])
+        if K.int_shape(layer.weights[0]) != weights[0].shape:
+            weights[0] = np.transpose(weights[0], (3, 2, 0, 1))
+            if layer.__class__.__name__ == 'ConvLSTM2D':
+                weights[1] = np.transpose(weights[1], (3, 2, 0, 1))
     return weights
 
 

keras/engine/training.py

@@ -50,6 +50,8 @@ def _standardize_input_data(data, names, shapes=None,
     # Raises
         ValueError: in case of improperly formatted user-provided data.
     """
+    if not names:
+        return []
     if data is None:
         return [None for _ in range(len(names))]
     if isinstance(data, dict):
@@ -63,7 +65,8 @@ def _standardize_input_data(data, names, shapes=None,
     elif isinstance(data, list):
         if len(data) != len(names):
             if data and hasattr(data[0], 'shape'):
-                raise ValueError('Error when checking ' + exception_prefix +
+                raise ValueError('Error when checking model ' +
+                                 exception_prefix +
                                  ': the list of Numpy arrays '
                                  'that you are passing to your model '
                                  'is not the size the model expected. '
@@ -77,7 +80,8 @@ def _standardize_input_data(data, names, shapes=None,
                     data = [np.asarray(data)]
                 else:
                     raise ValueError(
-                        'Error when checking ' + exception_prefix +
+                        'Error when checking model ' +
+                        exception_prefix +
                         ': you are passing a list as '
                         'input to your model, '
                         'but the model expects '
@@ -88,15 +92,17 @@ def _standardize_input_data(data, names, shapes=None,
         arrays = data
     else:
         if not hasattr(data, 'shape'):
-            raise TypeError('Error when checking ' + exception_prefix +
+            raise TypeError('Error when checking model ' +
+                            exception_prefix +
                             ': data should be a Numpy array, '
                             'or list/dict of Numpy arrays. '
                             'Found: ' + str(data)[:200] + '...')
-        if len(names) != 1:
+        if len(names) > 1:
             # Case: model expects multiple inputs but only received
             # a single Numpy array.
             raise ValueError('The model expects ' + str(len(names)) +
-                             ' input arrays, but only received one array. '
+                             exception_prefix +
+                             ' arrays, but only received one array. '
                              'Found: array with shape ' + str(data.shape))
         arrays = [data]
 
@@ -235,7 +241,7 @@ def _check_array_lengths(inputs, targets, weights):
 
 
 def _check_loss_and_target_compatibility(targets, loss_fns, output_shapes):
-    """Does validation on the compatiblity of targets and loss functions.
+    """Does validation on the compatibility of targets and loss functions.
 
     This helps prevent users from using loss functions incorrectly.
 
@@ -679,6 +685,8 @@ class Model(Container):
                 See [losses](/losses).
                 If the model has multiple outputs, you can use a different loss
                 on each output by passing a dictionary or a list of losses.
+                The loss value that will be minimized by the model
+                will then be the sum of all individual losses.
             metrics: list of metrics to be evaluated by the model
                 during training and testing.
                 Typically you will use `metrics=['accuracy']`.
@@ -688,6 +696,9 @@ class Model(Container):
             loss_weights: Optional list or dictionary specifying scalar
                 coefficients (Python floats) to weight the loss contributions
                 of different model outputs.
+                The loss value that will be minimized by the model
+                will then be the *weighted sum* of all individual losses,
+                weighted by the `loss_weights` coefficients.
                 If a list, it is expected to have a 1:1 mapping
                 to the model's outputs. If a tensor, it is expected to map
                 output names (strings) to scalar coefficients.
@@ -698,7 +709,8 @@ class Model(Container):
                 `sample_weight_mode` on each output by passing a
                 dictionary or a list of modes.
             **kwargs: when using the Theano backend, these arguments
-                are passed into K.function. Ignored for Tensorflow backend.
+                are passed into K.function. When using the Tensorflow backend,
+                these arguments are passed into `tf.Session.run`.
 
         # Raises
             ValueError: In case of invalid arguments for
@@ -1005,6 +1017,7 @@ class Model(Container):
             self.train_function = K.function(inputs,
                                              [self.total_loss] + self.metrics_tensors,
                                              updates=updates,
+                                             name='train_function',
                                              **self._function_kwargs)
 
     def _make_test_function(self):
@@ -1019,6 +1032,7 @@ class Model(Container):
             self.test_function = K.function(inputs,
                                             [self.total_loss] + self.metrics_tensors,
                                             updates=self.state_updates,
+                                            name='test_function',
                                             **self._function_kwargs)
 
     def _make_predict_function(self):
@@ -1035,6 +1049,7 @@ class Model(Container):
             self.predict_function = K.function(inputs,
                                                self.outputs,
                                                updates=self.state_updates,
+                                               name='predict_function',
                                                **kwargs)
 
     def _fit_loop(self, f, ins, out_labels=None, batch_size=32,
@@ -1126,7 +1141,7 @@ class Model(Container):
                 batch_ids = index_array[batch_start:batch_end]
                 try:
                     if isinstance(ins[-1], float):
-                        # do not slice the training phase flag
+                        # Do not slice the training phase flag.
                         ins_batch = _slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
                     else:
                         ins_batch = _slice_arrays(ins, batch_ids)
@@ -1145,17 +1160,17 @@ class Model(Container):
                     batch_logs[l] = o
 
                 callbacks.on_batch_end(batch_index, batch_logs)
+                if callback_model.stop_training:
+                    break
 
-                if batch_index == len(batches) - 1:  # last batch
-                    # validation
+                if batch_index == len(batches) - 1:  # Last batch.
                     if do_validation:
-                        # replace with self._evaluate
                         val_outs = self._test_loop(val_f, val_ins,
                                                    batch_size=batch_size,
                                                    verbose=0)
                         if not isinstance(val_outs, list):
                             val_outs = [val_outs]
-                        # same labels assumed
+                        # Same labels assumed.
                         for l, o in zip(out_labels, val_outs):
                             epoch_logs['val_' + l] = o
             callbacks.on_epoch_end(epoch, epoch_logs)
@@ -1195,7 +1210,7 @@ class Model(Container):
         for batch_index, (batch_start, batch_end) in enumerate(batches):
             batch_ids = index_array[batch_start:batch_end]
             if ins and isinstance(ins[-1], float):
-                # do not slice the training phase flag
+                # Do not slice the training phase flag.
                 ins_batch = _slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
             else:
                 ins_batch = _slice_arrays(ins, batch_ids)
@@ -1249,7 +1264,7 @@ class Model(Container):
         for batch_index, (batch_start, batch_end) in enumerate(batches):
             batch_ids = index_array[batch_start:batch_end]
             if isinstance(ins[-1], float):
-                # do not slice the training phase flag
+                # Do not slice the training phase flag.
                 ins_batch = _slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
             else:
                 ins_batch = _slice_arrays(ins, batch_ids)
@@ -1293,11 +1308,11 @@ class Model(Container):
         x = _standardize_input_data(x, self._feed_input_names,
                                     self._feed_input_shapes,
                                     check_batch_axis=False,
-                                    exception_prefix='model input')
+                                    exception_prefix='input')
         y = _standardize_input_data(y, self._feed_output_names,
                                     output_shapes,
                                     check_batch_axis=False,
-                                    exception_prefix='model target')
+                                    exception_prefix='target')
         sample_weights = _standardize_sample_weights(sample_weight,
                                                      self._feed_output_names)
         class_weights = _standardize_class_weights(class_weight,
@@ -1318,6 +1333,20 @@ class Model(Container):
                                  str(x[0].shape[0]) + ' samples')
         return x, y, sample_weights
 
+    def _get_deduped_metrics_names(self):
+        out_labels = self.metrics_names
+
+        # Rename duplicated metrics name
+        # (can happen with an output layer shared among multiple dataflows).
+        deduped_out_labels = []
+        for i, label in enumerate(out_labels):
+            new_label = label
+            if out_labels.count(label) > 1:
+                dup_idx = out_labels[:i].count(label)
+                new_label += '_' + str(dup_idx + 1)
+            deduped_out_labels.append(new_label)
+        return deduped_out_labels
+
     def fit(self, x=None,
             y=None,
             batch_size=32,
@@ -1347,7 +1376,7 @@ class Model(Container):
             batch_size: integer. Number of samples per gradient update.
             epochs: integer, the number of times to iterate
                 over the training data arrays.
-                verbose: 0, 1, or 2. Verbosity mode.
+            verbose: 0, 1, or 2. Verbosity mode.
                 0 = silent, 1 = verbose, 2 = one log line per epoch.
             callbacks: list of callbacks to be called during training.
                 See [callbacks](/callbacks).
@@ -1397,14 +1426,14 @@ class Model(Container):
         if kwargs:
             raise TypeError('Unrecognized keyword arguments: ' + str(kwargs))
 
-        # validate user data
+        # Validate user data.
         x, y, sample_weights = self._standardize_user_data(
             x, y,
             sample_weight=sample_weight,
             class_weight=class_weight,
             check_batch_axis=False,
             batch_size=batch_size)
-        # prepare validation data
+        # Prepare validation data.
         if validation_data:
             do_validation = True
             if len(validation_data) == 2:
@@ -1433,7 +1462,10 @@ class Model(Container):
 
         elif validation_split and 0. < validation_split < 1.:
             do_validation = True
-            split_at = int(len(x[0]) * (1. - validation_split))
+            if hasattr(x[0], 'shape'):
+                split_at = int(x[0].shape[0] * (1. - validation_split))
+            else:
+                split_at = int(len(x[0]) * (1. - validation_split))
             x, val_x = (_slice_arrays(x, 0, split_at), _slice_arrays(x, split_at))
             y, val_y = (_slice_arrays(y, 0, split_at), _slice_arrays(y, split_at))
             sample_weights, val_sample_weights = (
@@ -1450,7 +1482,7 @@ class Model(Container):
             val_f = None
             val_ins = None
 
-        # prepare input arrays and training function
+        # Prepare input arrays and training function.
         if self.uses_learning_phase and not isinstance(K.learning_phase(), int):
             ins = x + y + sample_weights + [1.]
         else:
@@ -1458,26 +1490,15 @@ class Model(Container):
         self._make_train_function()
         f = self.train_function
 
-        # prepare display labels
-        out_labels = self.metrics_names
-
-        # rename duplicated metrics name
-        # (can happen with an output layer shared among multiple dataflows)
-        deduped_out_labels = []
-        for i, label in enumerate(out_labels):
-            new_label = label
-            if out_labels.count(label) > 1:
-                dup_idx = out_labels[:i].count(label)
-                new_label += '_' + str(dup_idx + 1)
-            deduped_out_labels.append(new_label)
-        out_labels = deduped_out_labels
+        # Prepare display labels.
+        out_labels = self._get_deduped_metrics_names()
 
         if do_validation:
             callback_metrics = copy.copy(out_labels) + ['val_' + n for n in out_labels]
         else:
             callback_metrics = copy.copy(out_labels)
 
-        # delegate logic to _fit_loop
+        # Delegate logic to `_fit_loop`.
         return self._fit_loop(f, ins, out_labels=out_labels,
                               batch_size=batch_size, epochs=epochs,
                               verbose=verbose, callbacks=callbacks,
@@ -1512,13 +1533,13 @@ class Model(Container):
             and/or metrics). The attribute `model.metrics_names` will give you
             the display labels for the scalar outputs.
         """
-        # validate user data
+        # Validate user data.
         x, y, sample_weights = self._standardize_user_data(
             x, y,
             sample_weight=sample_weight,
             check_batch_axis=False,
             batch_size=batch_size)
-        # prepare inputs, delegate logic to _test_loop
+        # Prepare inputs, delegate logic to `_test_loop`.
         if self.uses_learning_phase and not isinstance(K.learning_phase(), int):
             ins = x + y + sample_weights + [0.]
         else:
@@ -1549,7 +1570,7 @@ class Model(Container):
                 or in case a stateful model receives a number of samples
                 that is not a multiple of the batch size.
         """
-        # validate user data
+        # Validate user data.
         x = _standardize_input_data(x, self._feed_input_names,
                                     self._feed_input_shapes,
                                     check_batch_axis=False)
@@ -1562,7 +1583,7 @@ class Model(Container):
                                  str(x[0].shape[0]) + ' samples. '
                                  'Batch size: ' + str(batch_size) + '.')
 
-        # prepare inputs, delegate logic to _predict_loop
+        # Prepare inputs, delegate logic to `_predict_loop`.
         if self.uses_learning_phase and not isinstance(K.learning_phase(), int):
             ins = x + [0.]
         else:
@@ -1713,7 +1734,7 @@ class Model(Container):
                 All arrays should contain the same number of samples.
                 The generator is expected to loop over its data
                 indefinitely. An epoch finishes when `steps_per_epoch`
-                samples have been seen by the model.
+                batches have been seen by the model.
             steps_per_epoch: Total number of steps (batches of samples)
                 to yield from `generator` before declaring one epoch
                 finished and starting the next epoch. It should typically
@@ -1785,7 +1806,8 @@ class Model(Container):
                              'you must specify a value for '
                              '`validation_steps`.')
 
-        out_labels = self.metrics_names
+        # Prepare display labels.
+        out_labels = self._get_deduped_metrics_names()
         callback_metrics = out_labels + ['val_' + n for n in out_labels]
 
         # prepare callbacks
@@ -1823,8 +1845,11 @@ class Model(Container):
                                  str(validation_data))
             val_x, val_y, val_sample_weights = self._standardize_user_data(
                 val_x, val_y, val_sample_weight)
+            val_data = val_x + val_y + val_sample_weights
+            if self.uses_learning_phase and not isinstance(K.learning_phase(), int):
+                val_data += [0.]
             for cbk in callbacks:
-                cbk.validation_data = val_x + [val_y, val_sample_weights]
+                cbk.validation_data = val_data
         enqueuer = None
 
         try:
@@ -1931,7 +1956,7 @@ class Model(Container):
         The generator should return the same kind of data
         as accepted by `test_on_batch`.
 
-        Arguments:
+        # Arguments
             generator: Generator yielding tuples (inputs, targets)
                 or (inputs, targets, sample_weights)
             steps: Total number of steps (batches of samples)

keras/layers/__init__.py

@@ -21,6 +21,14 @@ from ..legacy.layers import *
 
 
 def serialize(layer):
+    """Serialize a layer.
+
+    # Arguments
+        layer: a Layer object.
+
+    # Returns
+        dictionary with config.
+    """
     return {'class_name': layer.__class__.__name__,
             'config': layer.get_config()}
 

keras/layers/advanced_activations.py

@@ -104,7 +104,7 @@ class PReLU(Layer):
             for i in self.shared_axes:
                 param_shape[i - 1] = 1
                 self.param_broadcast[i - 1] = True
-        self.alpha = self.add_weight(param_shape,
+        self.alpha = self.add_weight(shape=param_shape,
                                      name='alpha',
                                      initializer=self.alpha_initializer,
                                      regularizer=self.alpha_regularizer,
@@ -202,7 +202,7 @@ class ThresholdedReLU(Layer):
         self.theta = K.cast_to_floatx(theta)
 
     def call(self, inputs, mask=None):
-        return inputs * K.cast(inputs > self.theta, K.floatx())
+        return inputs * K.cast(K.greater(inputs, self.theta), K.floatx())
 
     def get_config(self):
         config = {'theta': float(self.theta)}

keras/layers/convolutional.py

@@ -127,13 +127,13 @@ class _Conv(Layer):
         input_dim = input_shape[channel_axis]
         kernel_shape = self.kernel_size + (input_dim, self.filters)
 
-        self.kernel = self.add_weight(kernel_shape,
+        self.kernel = self.add_weight(shape=kernel_shape,
                                       initializer=self.kernel_initializer,
                                       name='kernel',
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         if self.use_bias:
-            self.bias = self.add_weight((self.filters,),
+            self.bias = self.add_weight(shape=(self.filters,),
                                         initializer=self.bias_initializer,
                                         name='bias',
                                         regularizer=self.bias_regularizer,
@@ -256,6 +256,9 @@ class Conv1D(_Conv):
             Specifying any stride value != 1 is incompatible with specifying
             any `dilation_rate` value != 1.
         padding: One of `"valid"`, `"causal"` or `"same"` (case-insensitive).
+            `"valid"` means "no padding".
+            `"same"` results in padding the input such that
+            the output has the same length as the original input.
             `"causal"` results in causal (dilated) convolutions, e.g. output[t]
             does not depend on input[t+1:]. Useful when modeling temporal data
             where the model should not violate the temporal order.
@@ -721,13 +724,13 @@ class Conv2DTranspose(Conv2D):
         input_dim = input_shape[channel_axis]
         kernel_shape = self.kernel_size + (self.filters, input_dim)
 
-        self.kernel = self.add_weight(kernel_shape,
+        self.kernel = self.add_weight(shape=kernel_shape,
                                       initializer=self.kernel_initializer,
                                       name='kernel',
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         if self.use_bias:
-            self.bias = self.add_weight((self.filters,),
+            self.bias = self.add_weight(shape=(self.filters,),
                                         initializer=self.bias_initializer,
                                         name='bias',
                                         regularizer=self.bias_regularizer,
@@ -952,20 +955,20 @@ class SeparableConv2D(Conv2D):
                                   self.filters)
 
         self.depthwise_kernel = self.add_weight(
-            depthwise_kernel_shape,
+            shape=depthwise_kernel_shape,
             initializer=self.depthwise_initializer,
             name='depthwise_kernel',
             regularizer=self.depthwise_regularizer,
             constraint=self.depthwise_constraint)
         self.pointwise_kernel = self.add_weight(
-            pointwise_kernel_shape,
+            shape=pointwise_kernel_shape,
             initializer=self.pointwise_initializer,
             name='pointwise_kernel',
             regularizer=self.pointwise_regularizer,
             constraint=self.pointwise_constraint)
 
         if self.use_bias:
-            self.bias = self.add_weight((self.filters,),
+            self.bias = self.add_weight(shape=(self.filters,),
                                         initializer=self.bias_initializer,
                                         name='bias',
                                         regularizer=self.bias_regularizer,
@@ -1252,7 +1255,7 @@ class ZeroPadding1D(Layer):
 class ZeroPadding2D(Layer):
     """Zero-padding layer for 2D input (e.g. picture).
 
-    This layer can add rows and columns or zeros
+    This layer can add rows and columns of zeros
     at the top, bottom, left and right side of an image tensor.
 
     # Arguments
@@ -1430,15 +1433,15 @@ class ZeroPadding3D(Layer):
     def compute_output_shape(self, input_shape):
         if self.data_format == 'channels_first':
             if input_shape[2] is not None:
-                dim1 = input_shape[2] + 2 * self.padding[0][0]
+                dim1 = input_shape[2] + self.padding[0][0] + self.padding[0][1]
             else:
                 dim1 = None
             if input_shape[3] is not None:
-                dim2 = input_shape[3] + 2 * self.padding[1][0]
+                dim2 = input_shape[3] + self.padding[1][0] + self.padding[1][1]
             else:
                 dim2 = None
             if input_shape[4] is not None:
-                dim3 = input_shape[4] + 2 * self.padding[2][0]
+                dim3 = input_shape[4] + self.padding[2][0] + self.padding[2][1]
             else:
                 dim3 = None
             return (input_shape[0],
@@ -1448,15 +1451,15 @@ class ZeroPadding3D(Layer):
                     dim3)
         elif self.data_format == 'channels_last':
             if input_shape[1] is not None:
-                dim1 = input_shape[1] + 2 * self.padding[0][1]
+                dim1 = input_shape[1] + self.padding[0][0] + self.padding[0][1]
             else:
                 dim1 = None
             if input_shape[2] is not None:
-                dim2 = input_shape[2] + 2 * self.padding[1][1]
+                dim2 = input_shape[2] + self.padding[1][0] + self.padding[1][1]
             else:
                 dim2 = None
             if input_shape[3] is not None:
-                dim3 = input_shape[3] + 2 * self.padding[2][1]
+                dim3 = input_shape[3] + self.padding[2][0] + self.padding[2][1]
             else:
                 dim3 = None
             return (input_shape[0],
@@ -1571,7 +1574,7 @@ class Cropping2D(Layer):
         model.add(Cropping2D(cropping=((2, 2), (4, 4)),
                              input_shape=(28, 28, 3)))
         # now model.output_shape == (None, 24, 20, 3)
-        model.add(Conv2D(64, (3, 3), padding='same))
+        model.add(Conv2D(64, (3, 3), padding='same'))
         model.add(Cropping2D(cropping=((2, 2), (2, 2))))
         # now model.output_shape == (None, 20, 16. 64)
     ```

keras/layers/convolutional_recurrent.py

@@ -105,9 +105,12 @@ class ConvRecurrent2D(Recurrent):
         self.return_sequences = return_sequences
         self.go_backwards = go_backwards
         self.stateful = stateful
-        self.input_spec = InputSpec(ndim=5)
+        self.input_spec = [InputSpec(ndim=5)]
+        self.state_spec = None
 
     def compute_output_shape(self, input_shape):
+        if isinstance(input_shape, list):
+            input_shape = input_shape[0]
         if self.data_format == 'channels_first':
             rows = input_shape[3]
             cols = input_shape[4]
@@ -328,11 +331,13 @@ class ConvLSTM2D(ConvRecurrent2D):
 
         self.dropout = min(1., max(0., dropout))
         self.recurrent_dropout = min(1., max(0., recurrent_dropout))
+        self.state_spec = [InputSpec(ndim=4), InputSpec(ndim=4)]
 
     def build(self, input_shape):
-        # TODO: better handling of input spec
-        self.input_spec = InputSpec(shape=input_shape)
-
+        if isinstance(input_shape, list):
+            input_shape = input_shape[0]
+        batch_size = input_shape[0] if self.stateful else None
+        self.input_spec[0] = InputSpec(shape=(batch_size, None) + input_shape[2:])
         if self.stateful:
             self.reset_states()
         else:
@@ -347,23 +352,27 @@ class ConvLSTM2D(ConvRecurrent2D):
             raise ValueError('The channel dimension of the inputs '
                              'should be defined. Found `None`.')
         input_dim = input_shape[channel_axis]
+        state_shape = [None] * 4
+        state_shape[channel_axis] = input_dim
+        state_shape = tuple(state_shape)
+        self.state_spec = [InputSpec(shape=state_shape), InputSpec(shape=state_shape)]
         kernel_shape = self.kernel_size + (input_dim, self.filters * 4)
         self.kernel_shape = kernel_shape
         recurrent_kernel_shape = self.kernel_size + (self.filters, self.filters * 4)
 
-        self.kernel = self.add_weight(kernel_shape,
+        self.kernel = self.add_weight(shape=kernel_shape,
                                       initializer=self.kernel_initializer,
                                       name='kernel',
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         self.recurrent_kernel = self.add_weight(
-            recurrent_kernel_shape,
+            shape=recurrent_kernel_shape,
             initializer=self.recurrent_initializer,
             name='recurrent_kernel',
             regularizer=self.recurrent_regularizer,
             constraint=self.recurrent_constraint)
         if self.use_bias:
-            self.bias = self.add_weight((self.filters * 4,),
+            self.bias = self.add_weight(shape=(self.filters * 4,),
                                         initializer=self.bias_initializer,
                                         name='bias',
                                         regularizer=self.bias_regularizer,
@@ -396,7 +405,7 @@ class ConvLSTM2D(ConvRecurrent2D):
             self.bias_o = None
         self.built = True
 
-    def get_initial_states(self, inputs):
+    def get_initial_state(self, inputs):
         # (samples, timesteps, rows, cols, filters)
         initial_state = K.zeros_like(inputs)
         # (samples, rows, cols, filters)
@@ -413,7 +422,7 @@ class ConvLSTM2D(ConvRecurrent2D):
     def reset_states(self):
         if not self.stateful:
             raise RuntimeError('Layer must be stateful.')
-        input_shape = self.input_spec.shape
+        input_shape = self.input_spec[0].shape
         output_shape = self.compute_output_shape(input_shape)
         if not input_shape[0]:
             raise ValueError('If a RNN is stateful, a complete '

keras/layers/core.py

@@ -299,13 +299,13 @@ class Reshape(Layer):
     """Reshapes an output to a certain shape.
 
     # Arguments
-        target_shape: target shape. Tuple of integers,
-            does not include the samples dimension (batch size).
+        target_shape: target shape. Tuple of integers.
+            Does not include the batch axis.
 
     # Input shape
         Arbitrary, although all dimensions in the input shaped must be fixed.
         Use the keyword argument `input_shape`
-        (tuple of integers, does not include the samples axis)
+        (tuple of integers, does not include the batch axis)
         when using this layer as the first layer in a model.
 
     # Output shape
@@ -335,27 +335,22 @@ class Reshape(Layer):
         self.target_shape = tuple(target_shape)
 
     def _fix_unknown_dimension(self, input_shape, output_shape):
-        """Find and replace a missing dimension in an output shape.
+        """Finds and replaces a missing dimension in an output shape.
 
         This is a near direct port of the internal Numpy function
         `_fix_unknown_dimension` in `numpy/core/src/multiarray/shape.c`
 
         # Arguments
-            input_shape: shape of array being reshaped
-            output_shape: desired shape of the array with at most
+            input_shape: original shape of array being reshaped
+            output_shape: target shape of the array, with at most
                 a single -1 which indicates a dimension that should be
                 derived from the input shape.
 
         # Returns
-            The new output shape with a -1 replaced with its computed value.
-
-            Raises a ValueError if the total array size of the output_shape is
-            different then the input_shape, or more then one unknown dimension
-            is specified.
+            The new output shape with a `-1` replaced with its computed value.
 
         # Raises
-            ValueError: in case of invalid values
-                for `input_shape` or `input_shape`.
+            ValueError: if `input_shape` and `output_shape` do not match.
         """
         output_shape = list(output_shape)
         msg = 'total size of new array must be unchanged'
@@ -386,13 +381,11 @@ class Reshape(Layer):
 
     def call(self, inputs):
         # In case the target shape is not fully defined,
-        # we need access to the shape of x.
-        # solution:
-        # 1) rely on x._keras_shape
-        # 2) fallback: K.int_shape
+        # we need access to the shape of `inputs`.
+        # solution: rely on `K.int_shape`.
         target_shape = self.target_shape
         if -1 in target_shape:
-            # target shape not fully defined
+            # Target shape not fully defined.
             input_shape = None
             try:
                 input_shape = K.int_shape(inputs)
@@ -720,6 +713,16 @@ class Lambda(Layer):
         else:
             output_shape = config['output_shape']
 
+        # If arguments were numpy array, they have been saved as
+        # list. We need to recover the ndarray
+        if 'arguments' in config:
+            for key in config['arguments']:
+                if isinstance(config['arguments'][key], dict):
+                    arg_dict = config['arguments'][key]
+                    if 'type' in arg_dict and arg_dict['type'] == 'ndarray':
+                        # Overwrite the argument with its numpy translation
+                        config['arguments'][key] = np.array(arg_dict['value'])
+
         config['function'] = function
         config['output_shape'] = output_shape
         return cls(**config)
@@ -820,13 +823,13 @@ class Dense(Layer):
         assert len(input_shape) >= 2
         input_dim = input_shape[-1]
 
-        self.kernel = self.add_weight((input_dim, self.units),
+        self.kernel = self.add_weight(shape=(input_dim, self.units),
                                       initializer=self.kernel_initializer,
                                       name='kernel',
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         if self.use_bias:
-            self.bias = self.add_weight((self.units,),
+            self.bias = self.add_weight(shape=(self.units,),
                                         initializer=self.bias_initializer,
                                         name='bias',
                                         regularizer=self.bias_regularizer,

keras/layers/embeddings.py

@@ -94,7 +94,7 @@ class Embedding(Layer):
 
     def build(self, input_shape):
         self.embeddings = self.add_weight(
-            (self.input_dim, self.output_dim),
+            shape=(self.input_dim, self.output_dim),
             initializer=self.embeddings_initializer,
             name='embeddings',
             regularizer=self.embeddings_regularizer,
@@ -108,11 +108,25 @@ class Embedding(Layer):
             return K.not_equal(inputs, 0)
 
     def compute_output_shape(self, input_shape):
-        if not self.input_length:
-            input_length = input_shape[1]
+        if self.input_length is None:
+            return input_shape + (self.output_dim,)
         else:
-            input_length = self.input_length
-        return (input_shape[0], input_length, self.output_dim)
+            # input_length can be tuple if input is 3D or higher
+            if isinstance(self.input_length, (list, tuple)):
+                in_lens = list(self.input_length)
+            else:
+                in_lens = [self.input_length]
+            if len(in_lens) != len(input_shape) - 1:
+                ValueError('"input_length" is %s, but received input has shape %s' %
+                           (str(self.input_length), str(input_shape)))
+            else:
+                for i, (s1, s2) in enumerate(zip(in_lens, input_shape[1:])):
+                    if s1 is not None and s2 is not None and s1 != s2:
+                        ValueError('"input_length" is %s, but received input has shape %s' %
+                                   (str(self.input_length), str(input_shape)))
+                    elif s1 is None:
+                        in_lens[i] = s2
+            return (input_shape[0],) + tuple(in_lens) + (self.output_dim,)
 
     def call(self, inputs):
         if K.dtype(inputs) != 'int32':

keras/layers/local.py

@@ -122,14 +122,14 @@ class LocallyConnected1D(Layer):
                              self.kernel_size[0] * input_dim,
                              self.filters)
         self.kernel = self.add_weight(
-            self.kernel_shape,
+            shape=self.kernel_shape,
             initializer=self.kernel_initializer,
             name='kernel',
             regularizer=self.kernel_regularizer,
             constraint=self.kernel_constraint)
         if self.use_bias:
             self.bias = self.add_weight(
-                (output_length, self.filters),
+                shape=(output_length, self.filters),
                 initializer=self.bias_initializer,
                 name='bias',
                 regularizer=self.bias_regularizer,
@@ -147,22 +147,11 @@ class LocallyConnected1D(Layer):
         return (input_shape[0], length, self.filters)
 
     def call(self, inputs):
-        stride = self.strides[0]
-        output_length, feature_dim, filters = self.kernel_shape
-
-        xs = []
-        for i in range(output_length):
-            slice_length = slice(i * stride,
-                                 i * stride + self.kernel_size[0])
-            xs.append(K.reshape(inputs[:, slice_length, :],
-                                (1, -1, feature_dim)))
-        x_aggregate = K.concatenate(xs, axis=0)
-        # Shape: `(output_length, batch_size, filters)`.
-        output = K.batch_dot(x_aggregate, self.kernel)
-        output = K.permute_dimensions(output, (1, 0, 2))
+        output_length, _, filters = self.kernel_shape
 
+        output = K.local_conv1d(inputs, self.kernel, self.kernel_size, self.strides)
         if self.use_bias:
-            output += K.reshape(self.bias, (1, output_length, filters))
+            output = K.bias_add(output, self.bias)
         if self.activation is not None:
             output = self.activation(output)
         return output
@@ -325,13 +314,13 @@ class LocallyConnected2D(Layer):
         self.kernel_shape = (output_row * output_col,
                              self.kernel_size[0] * self.kernel_size[1] * input_filter,
                              self.filters)
-        self.kernel = self.add_weight(self.kernel_shape,
+        self.kernel = self.add_weight(shape=self.kernel_shape,
                                       initializer=self.kernel_initializer,
                                       name='kernel',
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         if self.use_bias:
-            self.bias = self.add_weight((output_row, output_col, self.filters),
+            self.bias = self.add_weight(shape=(output_row, output_col, self.filters),
                                         initializer=self.bias_initializer,
                                         name='bias',
                                         regularizer=self.bias_regularizer,
@@ -363,62 +352,19 @@ class LocallyConnected2D(Layer):
             return (input_shape[0], rows, cols, self.filters)
 
     def call(self, inputs):
-        stride_row, stride_col = self.strides
-        _, feature_dim, filters = self.kernel_shape
+        _, _, filters = self.kernel_shape
 
-        if self.data_format == 'channels_first':
-            if K.backend() == 'theano':
-                output = []
-                for i in range(self.output_row):
-                    for j in range(self.output_col):
-                        slice_row = slice(i * stride_row,
-                                          i * stride_row + self.kernel_size[0])
-                        slice_col = slice(j * stride_col,
-                                          j * stride_col + self.kernel_size[1])
-                        x_flatten = K.reshape(inputs[:, :, slice_row, slice_col],
-                                              (1, -1, feature_dim))
-                        output.append(K.dot(x_flatten,
-                                      self.kernel[i * self.output_col + j, :, :]))
-                output = K.concatenate(output, axis=0)
-            else:
-                xs = []
-                for i in range(self.output_row):
-                    for j in range(self.output_col):
-                        slice_row = slice(i * stride_row,
-                                          i * stride_row + self.kernel_size[0])
-                        slice_col = slice(j * stride_col,
-                                          j * stride_col + self.kernel_size[1])
-                        xs.append(K.reshape(inputs[:, :, slice_row, slice_col],
-                                            (1, -1, feature_dim)))
-                x_aggregate = K.concatenate(xs, axis=0)
-                output = K.batch_dot(x_aggregate, self.kernel)
-            output = K.reshape(output,
-                               (self.output_row, self.output_col, -1, filters))
-            output = K.permute_dimensions(output, (2, 3, 0, 1))
-
-        elif self.data_format == 'channels_last':
-            xs = []
-            for i in range(self.output_row):
-                for j in range(self.output_col):
-                    slice_row = slice(i * stride_row,
-                                      i * stride_row + self.kernel_size[0])
-                    slice_col = slice(j * stride_col,
-                                      j * stride_col + self.kernel_size[1])
-                    xs.append(K.reshape(inputs[:, slice_row, slice_col, :],
-                                        (1, -1, feature_dim)))
-            x_aggregate = K.concatenate(xs, axis=0)
-            output = K.batch_dot(x_aggregate, self.kernel)
-            output = K.reshape(output,
-                               (self.output_row, self.output_col, -1, filters))
-            output = K.permute_dimensions(output, (2, 0, 1, 3))
+        output = K.local_conv2d(inputs,
+                                self.kernel,
+                                self.kernel_size,
+                                self.strides,
+                                (self.output_row, self.output_col),
+                                self.data_format)
 
         if self.use_bias:
-            if self.data_format == 'channels_first':
-                output += K.reshape(self.bias,
-                                    (1, filters, self.output_row, self.output_col))
-            elif self.data_format == 'channels_last':
-                output += K.reshape(self.bias,
-                                    (1, self.output_row, self.output_col, filters))
+            if self.data_format == 'channels_first' or self.data_format == 'channels_last':
+                output = K.bias_add(output, self.bias, data_format=self.data_format)
+
         output = self.activation(output)
         return output
 

keras/layers/normalization.py

@@ -96,7 +96,7 @@ class BatchNormalization(Layer):
         shape = (dim,)
 
         if self.scale:
-            self.gamma = self.add_weight(shape,
+            self.gamma = self.add_weight(shape=shape,
                                          name='gamma',
                                          initializer=self.gamma_initializer,
                                          regularizer=self.gamma_regularizer,
@@ -104,7 +104,7 @@ class BatchNormalization(Layer):
         else:
             self.gamma = None
         if self.center:
-            self.beta = self.add_weight(shape,
+            self.beta = self.add_weight(shape=shape,
                                         name='beta',
                                         initializer=self.beta_initializer,
                                         regularizer=self.beta_regularizer,
@@ -112,12 +112,12 @@ class BatchNormalization(Layer):
         else:
             self.beta = None
         self.moving_mean = self.add_weight(
-            shape,
+            shape=shape,
             name='moving_mean',
             initializer=self.moving_mean_initializer,
             trainable=False)
         self.moving_variance = self.add_weight(
-            shape,
+            shape=shape,
             name='moving_variance',
             initializer=self.moving_variance_initializer,
             trainable=False)
@@ -135,55 +135,57 @@ class BatchNormalization(Layer):
         # Determines whether broadcasting is needed.
         needs_broadcasting = (sorted(reduction_axes) != list(range(ndim))[:-1])
 
-        normed, mean, variance = K.normalize_batch_in_training(
+        def normalize_inference():
+            if needs_broadcasting:
+                # In this case we must explictly broadcast all parameters.
+                broadcast_moving_mean = K.reshape(self.moving_mean,
+                                                  broadcast_shape)
+                broadcast_moving_variance = K.reshape(self.moving_variance,
+                                                      broadcast_shape)
+                if self.center:
+                    broadcast_beta = K.reshape(self.beta, broadcast_shape)
+                else:
+                    broadcast_beta = None
+                if self.scale:
+                    broadcast_gamma = K.reshape(self.gamma,
+                                                broadcast_shape)
+                else:
+                    broadcast_gamma = None
+                return K.batch_normalization(
+                    inputs,
+                    broadcast_moving_mean,
+                    broadcast_moving_variance,
+                    broadcast_beta,
+                    broadcast_gamma,
+                    epsilon=self.epsilon)
+            else:
+                return K.batch_normalization(
+                    inputs,
+                    self.moving_mean,
+                    self.moving_variance,
+                    self.beta,
+                    self.gamma,
+                    epsilon=self.epsilon)
+
+        # If the learning phase is *static* and set to inference:
+        if training in {0, False}:
+            return normalize_inference()
+
+        # If the learning is either dynamic, or set to training:
+        normed_training, mean, variance = K.normalize_batch_in_training(
             inputs, self.gamma, self.beta, reduction_axes,
             epsilon=self.epsilon)
 
-        if training in {0, False}:
-            return normed
-        else:
-            self.add_update([K.moving_average_update(self.moving_mean,
-                                                     mean,
-                                                     self.momentum),
-                             K.moving_average_update(self.moving_variance,
-                                                     variance,
-                                                     self.momentum)],
-                            inputs)
-
-            def normalize_inference():
-                if needs_broadcasting:
-                    # In this case we must explictly broadcast all parameters.
-                    broadcast_moving_mean = K.reshape(self.moving_mean,
-                                                      broadcast_shape)
-                    broadcast_moving_variance = K.reshape(self.moving_variance,
-                                                          broadcast_shape)
-                    if self.center:
-                        broadcast_beta = K.reshape(self.beta, broadcast_shape)
-                    else:
-                        broadcast_beta = None
-                    if self.scale:
-                        broadcast_gamma = K.reshape(self.gamma,
-                                                    broadcast_shape)
-                    else:
-                        broadcast_gamma = None
-                    return K.batch_normalization(
-                        inputs,
-                        broadcast_moving_mean,
-                        broadcast_moving_variance,
-                        broadcast_beta,
-                        broadcast_gamma,
-                        epsilon=self.epsilon)
-                else:
-                    return K.batch_normalization(
-                        inputs,
-                        self.moving_mean,
-                        self.moving_variance,
-                        self.beta,
-                        self.gamma,
-                        epsilon=self.epsilon)
+        self.add_update([K.moving_average_update(self.moving_mean,
+                                                 mean,
+                                                 self.momentum),
+                         K.moving_average_update(self.moving_variance,
+                                                 variance,
+                                                 self.momentum)],
+                        inputs)
 
         # Pick the normalized form corresponding to the training phase.
-        return K.in_train_phase(normed,
+        return K.in_train_phase(normed_training,
                                 normalize_inference,
                                 training=training)
 

keras/layers/recurrent.py

@@ -96,6 +96,8 @@ class Recurrent(Layer):
             `[(input_dim, output_dim), (output_dim, output_dim), (output_dim,)]`.
         return_sequences: Boolean. Whether to return the last output
             in the output sequence, or the full sequence.
+        return_state: Boolean. Whether to return the last state
+            in addition to the output.
         go_backwards: Boolean (default False).
             If True, process the input sequence backwards and return the
             reversed sequence.
@@ -139,6 +141,9 @@ class Recurrent(Layer):
         (Optional) 2D tensors with shape `(batch_size, output_dim)`.
 
     # Output shape
+        - if `return_state`: a list of tensors. The first tensor is
+            the output. The remaining tensors are the last states,
+            each with shape `(batch_size, units)`.
         - if `return_sequences`: 3D tensor with shape
             `(batch_size, timesteps, units)`.
         - else, 2D tensor with shape `(batch_size, units)`.
@@ -170,14 +175,20 @@ class Recurrent(Layer):
         To reset the states of your model, call `.reset_states()` on either
         a specific layer, or on your entire model.
 
-    # Note on specifying initial states in RNNs
-        You can specify the initial state of RNN layers by calling them with
-        the keyword argument `initial_state`. The value of `initial_state`
-        should be a tensor or list of tensors representing the initial state
-        of the RNN layer.
+    # Note on specifying the initial state of RNNs
+        You can specify the initial state of RNN layers symbolically by
+        calling them with the keyword argument `initial_state`. The value of
+        `initial_state` should be a tensor or list of tensors representing
+        the initial state of the RNN layer.
+
+        You can specify the initial state of RNN layers numerically by
+        calling `reset_states` with the keyword argument `states`. The value of
+        `states` should be a numpy array or list of numpy arrays representing
+        the initial state of the RNN layer.
     """
 
     def __init__(self, return_sequences=False,
+                 return_state=False,
                  go_backwards=False,
                  stateful=False,
                  unroll=False,
@@ -185,12 +196,16 @@ class Recurrent(Layer):
                  **kwargs):
         super(Recurrent, self).__init__(**kwargs)
         self.return_sequences = return_sequences
+        self.return_state = return_state
         self.go_backwards = go_backwards
+        if K.backend() == 'cntk' and stateful:
+            raise ValueError('Stateful RNN is not currently supported with CNTK.')
+
         self.stateful = stateful
         self.unroll = unroll
         self.implementation = implementation
         self.supports_masking = True
-        self.input_spec = InputSpec(ndim=3)
+        self.input_spec = [InputSpec(ndim=3)]
         self.state_spec = None
         self.dropout = 0
         self.recurrent_dropout = 0
@@ -198,16 +213,27 @@ class Recurrent(Layer):
     def compute_output_shape(self, input_shape):
         if isinstance(input_shape, list):
             input_shape = input_shape[0]
+
         if self.return_sequences:
-            return (input_shape[0], input_shape[1], self.units)
+            output_shape = (input_shape[0], input_shape[1], self.units)
         else:
-            return (input_shape[0], self.units)
+            output_shape = (input_shape[0], self.units)
+
+        if self.return_state:
+            state_shape = [(input_shape[0], self.units) for _ in self.states]
+            return [output_shape] + state_shape
+        else:
+            return output_shape
 
     def compute_mask(self, inputs, mask):
-        if self.return_sequences:
-            return mask
+        if isinstance(mask, list):
+            mask = mask[0]
+        output_mask = mask if self.return_sequences else None
+        if self.return_state:
+            state_mask = [None for _ in self.states]
+            return [output_mask] + state_mask
         else:
-            return None
+            return output_mask
 
     def step(self, inputs, states):
         raise NotImplementedError
@@ -215,14 +241,14 @@ class Recurrent(Layer):
     def get_constants(self, inputs, training=None):
         return []
 
-    def get_initial_states(self, inputs):
+    def get_initial_state(self, inputs):
         # build an all-zero tensor of shape (samples, output_dim)
         initial_state = K.zeros_like(inputs)  # (samples, timesteps, input_dim)
         initial_state = K.sum(initial_state, axis=(1, 2))  # (samples,)
         initial_state = K.expand_dims(initial_state)  # (samples, 1)
         initial_state = K.tile(initial_state, [1, self.units])  # (samples, output_dim)
-        initial_states = [initial_state for _ in range(len(self.states))]
-        return initial_states
+        initial_state = [initial_state for _ in range(len(self.states))]
+        return initial_state
 
     def preprocess_input(self, inputs, training=None):
         return inputs
@@ -232,51 +258,63 @@ class Recurrent(Layer):
         # and if it a Keras tensor,
         # then add it to the inputs and temporarily
         # modify the input spec to include the state.
-        if initial_state is not None:
-            if hasattr(initial_state, '_keras_history'):
-                # Compute the full input spec, including state
-                input_spec = self.input_spec
-                state_spec = self.state_spec
-                if not isinstance(state_spec, list):
-                    state_spec = [state_spec]
-                self.input_spec = [input_spec] + state_spec
+        if initial_state is None:
+            return super(Recurrent, self).__call__(inputs, **kwargs)
 
-                # Compute the full inputs, including state
-                if not isinstance(initial_state, (list, tuple)):
-                    initial_state = [initial_state]
-                inputs = [inputs] + list(initial_state)
+        if not isinstance(initial_state, (list, tuple)):
+            initial_state = [initial_state]
 
-                # Perform the call
-                output = super(Recurrent, self).__call__(inputs, **kwargs)
+        is_keras_tensor = hasattr(initial_state[0], '_keras_history')
+        for tensor in initial_state:
+            if hasattr(tensor, '_keras_history') != is_keras_tensor:
+                raise ValueError('The initial state of an RNN layer cannot be'
+                                 ' specified with a mix of Keras tensors and'
+                                 ' non-Keras tensors')
 
-                # Restore original input spec
-                self.input_spec = input_spec
-                return output
-            else:
-                kwargs['initial_state'] = initial_state
-        return super(Recurrent, self).__call__(inputs, **kwargs)
+        if is_keras_tensor:
+            # Compute the full input spec, including state
+            input_spec = self.input_spec
+            state_spec = self.state_spec
+            if not isinstance(input_spec, list):
+                input_spec = [input_spec]
+            if not isinstance(state_spec, list):
+                state_spec = [state_spec]
+            self.input_spec = input_spec + state_spec
 
-    def call(self, inputs, mask=None, initial_state=None, training=None):
+            # Compute the full inputs, including state
+            inputs = [inputs] + list(initial_state)
+
+            # Perform the call
+            output = super(Recurrent, self).__call__(inputs, **kwargs)
+
+            # Restore original input spec
+            self.input_spec = input_spec
+            return output
+        else:
+            kwargs['initial_state'] = initial_state
+            return super(Recurrent, self).__call__(inputs, **kwargs)
+
+    def call(self, inputs, mask=None, training=None, initial_state=None):
         # input shape: `(samples, time (padded with zeros), input_dim)`
         # note that the .build() method of subclasses MUST define
         # self.input_spec and self.state_spec with complete input shapes.
-        if initial_state is not None:
-            if not isinstance(initial_state, (list, tuple)):
-                initial_states = [initial_state]
-            else:
-                initial_states = list(initial_state)
         if isinstance(inputs, list):
-            initial_states = inputs[1:]
+            initial_state = inputs[1:]
             inputs = inputs[0]
+        elif initial_state is not None:
+            pass
         elif self.stateful:
-            initial_states = self.states
+            initial_state = self.states
         else:
-            initial_states = self.get_initial_states(inputs)
+            initial_state = self.get_initial_state(inputs)
 
-        if len(initial_states) != len(self.states):
+        if isinstance(mask, list):
+            mask = mask[0]
+
+        if len(initial_state) != len(self.states):
             raise ValueError('Layer has ' + str(len(self.states)) +
                              ' states but was passed ' +
-                             str(len(initial_states)) +
+                             str(len(initial_state)) +
                              ' initial states.')
         input_shape = K.int_shape(inputs)
         if self.unroll and input_shape[1] is None:
@@ -295,7 +333,7 @@ class Recurrent(Layer):
         preprocessed_input = self.preprocess_input(inputs, training=None)
         last_output, outputs, states = K.rnn(self.step,
                                              preprocessed_input,
-                                             initial_states,
+                                             initial_state,
                                              go_backwards=self.go_backwards,
                                              mask=mask,
                                              constants=constants,
@@ -313,17 +351,23 @@ class Recurrent(Layer):
             outputs._uses_learning_phase = True
 
         if self.return_sequences:
-            return outputs
+            output = outputs
         else:
-            return last_output
+            output = last_output
 
-    def reset_states(self, states_value=None):
+        if self.return_state:
+            if not isinstance(states, (list, tuple)):
+                states = [states]
+            else:
+                states = list(states)
+            return [output] + states
+        else:
+            return output
+
+    def reset_states(self, states=None):
         if not self.stateful:
             raise AttributeError('Layer must be stateful.')
-        if not self.input_spec:
-            raise RuntimeError('Layer has never been called '
-                               'and thus has no states.')
-        batch_size = self.input_spec.shape[0]
+        batch_size = self.input_spec[0].shape[0]
         if not batch_size:
             raise ValueError('If a RNN is stateful, it needs to know '
                              'its batch size. Specify the batch size '
@@ -335,34 +379,34 @@ class Recurrent(Layer):
                              '- If using the functional API, specify '
                              'the time dimension by passing a '
                              '`batch_shape` argument to your Input layer.')
-        if states_value is not None:
-            if not isinstance(states_value, (list, tuple)):
-                states_value = [states_value]
-            if len(states_value) != len(self.states):
-                raise ValueError('The layer has ' + str(len(self.states)) +
-                                 ' states, but the `states_value` '
-                                 'argument passed '
-                                 'only has ' + str(len(states_value)) +
-                                 ' entries')
+        # initialize state if None
         if self.states[0] is None:
             self.states = [K.zeros((batch_size, self.units))
                            for _ in self.states]
-            if not states_value:
-                return
-        for i, state in enumerate(self.states):
-            if states_value:
-                value = states_value[i]
+        elif states is None:
+            for state in self.states:
+                K.set_value(state, np.zeros((batch_size, self.units)))
+        else:
+            if not isinstance(states, (list, tuple)):
+                states = [states]
+            if len(states) != len(self.states):
+                raise ValueError('Layer ' + self.name + ' expects ' +
+                                 str(len(self.states)) + ' states, '
+                                 'but it received ' + str(len(states)) +
+                                 ' state values. Input received: ' +
+                                 str(states))
+            for index, (value, state) in enumerate(zip(states, self.states)):
                 if value.shape != (batch_size, self.units):
-                    raise ValueError(
-                        'Expected state #' + str(i) +
-                        ' to have shape ' + str((batch_size, self.units)) +
-                        ' but got array with shape ' + str(value.shape))
-            else:
-                value = np.zeros((batch_size, self.units))
-            K.set_value(state, value)
+                    raise ValueError('State ' + str(index) +
+                                     ' is incompatible with layer ' +
+                                     self.name + ': expected shape=' +
+                                     str((batch_size, self.units)) +
+                                     ', found shape=' + str(value.shape))
+                K.set_value(state, value)
 
     def get_config(self):
         config = {'return_sequences': self.return_sequences,
+                  'return_state': self.return_state,
                   'go_backwards': self.go_backwards,
                   'stateful': self.stateful,
                   'unroll': self.unroll,
@@ -457,6 +501,7 @@ class SimpleRNN(Recurrent):
 
         self.dropout = min(1., max(0., dropout))
         self.recurrent_dropout = min(1., max(0., recurrent_dropout))
+        self.state_spec = InputSpec(shape=(None, self.units))
 
     def build(self, input_shape):
         if isinstance(input_shape, list):
@@ -464,26 +509,25 @@ class SimpleRNN(Recurrent):
 
         batch_size = input_shape[0] if self.stateful else None
         self.input_dim = input_shape[2]
-        self.input_spec = InputSpec(shape=(batch_size, None, self.input_dim))
-        self.state_spec = InputSpec(shape=(batch_size, self.units))
+        self.input_spec[0] = InputSpec(shape=(batch_size, None, self.input_dim))
 
         self.states = [None]
         if self.stateful:
             self.reset_states()
 
-        self.kernel = self.add_weight((self.input_dim, self.units),
+        self.kernel = self.add_weight(shape=(self.input_dim, self.units),
                                       name='kernel',
                                       initializer=self.kernel_initializer,
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         self.recurrent_kernel = self.add_weight(
-            (self.units, self.units),
+            shape=(self.units, self.units),
             name='recurrent_kernel',
             initializer=self.recurrent_initializer,
             regularizer=self.recurrent_regularizer,
             constraint=self.recurrent_constraint)
         if self.use_bias:
-            self.bias = self.add_weight((self.units,),
+            self.bias = self.add_weight(shape=(self.units,),
                                         name='bias',
                                         initializer=self.bias_initializer,
                                         regularizer=self.bias_regularizer,
@@ -676,6 +720,7 @@ class GRU(Recurrent):
 
         self.dropout = min(1., max(0., dropout))
         self.recurrent_dropout = min(1., max(0., recurrent_dropout))
+        self.state_spec = InputSpec(shape=(None, self.units))
 
     def build(self, input_shape):
         if isinstance(input_shape, list):
@@ -683,29 +728,28 @@ class GRU(Recurrent):
 
         batch_size = input_shape[0] if self.stateful else None
         self.input_dim = input_shape[2]
-        self.input_spec = InputSpec(shape=(batch_size, None, self.input_dim))
-        self.state_spec = InputSpec(shape=(batch_size, self.units))
+        self.input_spec[0] = InputSpec(shape=(batch_size, None, self.input_dim))
 
         self.states = [None]
         if self.stateful:
             self.reset_states()
 
-        self.kernel = self.add_weight((self.input_dim, self.units * 3),
+        self.kernel = self.add_weight(shape=(self.input_dim, self.units * 3),
                                       name='kernel',
                                       initializer=self.kernel_initializer,
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         self.recurrent_kernel = self.add_weight(
-            (self.units, self.units * 3),
+            shape=(self.units, self.units * 3),
             name='recurrent_kernel',
             initializer=self.recurrent_initializer,
             regularizer=self.recurrent_regularizer,
             constraint=self.recurrent_constraint)
 
         if self.use_bias:
-            self.bias = self.add_weight((self.units * 3,),
+            self.bias = self.add_weight(shape=(self.units * 3,),
                                         name='bias',
-                                        initializer='zero',
+                                        initializer=self.bias_initializer,
                                         regularizer=self.bias_regularizer,
                                         constraint=self.bias_constraint)
         else:
@@ -955,6 +999,8 @@ class LSTM(Recurrent):
 
         self.dropout = min(1., max(0., dropout))
         self.recurrent_dropout = min(1., max(0., recurrent_dropout))
+        self.state_spec = [InputSpec(shape=(None, self.units)),
+                           InputSpec(shape=(None, self.units))]
 
     def build(self, input_shape):
         if isinstance(input_shape, list):
@@ -962,36 +1008,39 @@ class LSTM(Recurrent):
 
         batch_size = input_shape[0] if self.stateful else None
         self.input_dim = input_shape[2]
-        self.input_spec = InputSpec(shape=(batch_size, None, self.input_dim))
-        self.state_spec = [InputSpec(shape=(batch_size, self.units)),
-                           InputSpec(shape=(batch_size, self.units))]
+        self.input_spec[0] = InputSpec(shape=(batch_size, None, self.input_dim))
 
         self.states = [None, None]
         if self.stateful:
             self.reset_states()
 
-        self.kernel = self.add_weight((self.input_dim, self.units * 4),
+        self.kernel = self.add_weight(shape=(self.input_dim, self.units * 4),
                                       name='kernel',
                                       initializer=self.kernel_initializer,
                                       regularizer=self.kernel_regularizer,
                                       constraint=self.kernel_constraint)
         self.recurrent_kernel = self.add_weight(
-            (self.units, self.units * 4),
+            shape=(self.units, self.units * 4),
             name='recurrent_kernel',
             initializer=self.recurrent_initializer,
             regularizer=self.recurrent_regularizer,
             constraint=self.recurrent_constraint)
 
         if self.use_bias:
-            self.bias = self.add_weight((self.units * 4,),
+            if self.unit_forget_bias:
+                def bias_initializer(shape, *args, **kwargs):
+                    return K.concatenate([
+                        self.bias_initializer((self.units,), *args, **kwargs),
+                        initializers.Ones()((self.units,), *args, **kwargs),
+                        self.bias_initializer((self.units * 2,), *args, **kwargs),
+                    ])
+            else:
+                bias_initializer = self.bias_initializer
+            self.bias = self.add_weight(shape=(self.units * 4,),
                                         name='bias',
-                                        initializer=self.bias_initializer,
+                                        initializer=bias_initializer,
                                         regularizer=self.bias_regularizer,
                                         constraint=self.bias_constraint)
-            if self.unit_forget_bias:
-                bias_value = np.zeros((self.units * 4,))
-                bias_value[self.units: self.units * 2] = 1.
-                K.set_value(self.bias, bias_value)
         else:
             self.bias = None
 

keras/legacy/interfaces.py

@@ -161,7 +161,7 @@ def recurrent_args_preprocessor(args, kwargs):
             kwargs.pop('forget_bias_init')
             warnings.warn('The `forget_bias_init` argument '
                           'has been ignored. Use `unit_forget_bias=True` '
-                          'instead to intialize with ones.', stacklevel=3)
+                          'instead to initialize with ones.', stacklevel=3)
     if 'input_dim' in kwargs:
         input_length = kwargs.pop('input_length', None)
         input_dim = kwargs.pop('input_dim')
@@ -461,7 +461,7 @@ def convlstm2d_args_preprocessor(args, kwargs):
         else:
             warnings.warn('The `forget_bias_init` argument '
                           'has been ignored. Use `unit_forget_bias=True` '
-                          'instead to intialize with ones.', stacklevel=3)
+                          'instead to initialize with ones.', stacklevel=3)
     args, kwargs, _converted = conv2d_args_preprocessor(args, kwargs)
     return args, kwargs, converted + _converted
 
@@ -602,3 +602,24 @@ legacy_model_constructor_support = generate_legacy_interface(
     allowed_positional_args=None,
     conversions=[('input', 'inputs'),
                  ('output', 'outputs')])
+
+legacy_input_support = generate_legacy_interface(
+    allowed_positional_args=None,
+    conversions=[('input_dtype', 'dtype')])
+
+
+def add_weight_args_preprocessing(args, kwargs):
+    if len(args) > 1:
+        if isinstance(args[1], (tuple, list)):
+            kwargs['shape'] = args[1]
+            args = (args[0],) + args[2:]
+            if len(args) > 1:
+                if isinstance(args[1], six.string_types):
+                    kwargs['name'] = args[1]
+                    args = (args[0],) + args[2:]
+    return args, kwargs, []
+
+
+legacy_add_weight_support = generate_legacy_interface(
+    allowed_positional_args=['name', 'shape'],
+    preprocessor=add_weight_args_preprocessing)

keras/losses.py

@@ -33,6 +33,18 @@ def hinge(y_true, y_pred):
     return K.mean(K.maximum(1. - y_true * y_pred, 0.), axis=-1)
 
 
+def categorical_hinge(y_true, y_pred):
+    pos = K.sum(y_true * y_pred, axis=-1)
+    neg = K.max((1. - y_true) * y_pred, axis=-1)
+    return K.maximum(0., neg - pos + 1.)
+
+
+def logcosh(y_true, y_pred):
+    def cosh(x):
+        return (K.exp(x) + K.exp(-x)) / 2
+    return K.mean(K.log(cosh(y_pred - y_true)), axis=-1)
+
+
 def categorical_crossentropy(y_true, y_pred):
     return K.categorical_crossentropy(y_pred, y_true)
 

keras/metrics.py

@@ -6,6 +6,7 @@ from .losses import mean_absolute_error
 from .losses import mean_absolute_percentage_error
 from .losses import mean_squared_logarithmic_error
 from .losses import hinge
+from .losses import logcosh
 from .losses import squared_hinge
 from .losses import categorical_crossentropy
 from .losses import sparse_categorical_crossentropy
@@ -35,6 +36,11 @@ def sparse_categorical_accuracy(y_true, y_pred):
 def top_k_categorical_accuracy(y_true, y_pred, k=5):
     return K.mean(K.in_top_k(y_pred, K.argmax(y_true, axis=-1), k), axis=-1)
 
+
+def sparse_top_k_categorical_accuracy(y_true, y_pred, k=5):
+    return K.mean(K.in_top_k(y_pred, K.cast(K.max(y_true, axis=-1), 'int32'), k), axis=-1)
+
+
 # Aliases
 
 mse = MSE = mean_squared_error

keras/models.py

@@ -74,7 +74,11 @@ def save_model(model, filepath, overwrite=True, include_optimizer=True):
 
         # if obj is any numpy type
         if type(obj).__module__ == np.__name__:
-            return obj.item()
+            if isinstance(obj, np.ndarray):
+                return {'type': type(obj),
+                        'value': obj.tolist()}
+            else:
+                return obj.item()
 
         # misc functions (e.g. loss function)
         if callable(obj):
@@ -140,8 +144,8 @@ def save_model(model, filepath, overwrite=True, include_optimizer=True):
                 weight_values = K.batch_get_value(symbolic_weights)
                 weight_names = []
                 for i, (w, val) in enumerate(zip(symbolic_weights, weight_values)):
-                    # Default values of symbolic_weights is /variable for theano
-                    if K.backend() == 'theano':
+                    # Default values of symbolic_weights is /variable for theano and cntk
+                    if K.backend() == 'theano' or K.backend() == 'cntk':
                         if hasattr(w, 'name') and w.name != "/variable":
                             name = str(w.name)
                         else:
@@ -167,7 +171,7 @@ def save_model(model, filepath, overwrite=True, include_optimizer=True):
     f.close()
 
 
-def load_model(filepath, custom_objects=None):
+def load_model(filepath, custom_objects=None, compile=True):
     """Loads a model saved via `save_model`.
 
     # Arguments
@@ -175,12 +179,16 @@ def load_model(filepath, custom_objects=None):
         custom_objects: Optional dictionary mapping names
             (strings) to custom classes or functions to be
             considered during deserialization.
+        compile: Boolean, whether to compile the model
+            after loading.
 
     # Returns
         A Keras model instance. If an optimizer was found
         as part of the saved model, the model is already
         compiled. Otherwise, the model is uncompiled and
-        a warning will be displayed.
+        a warning will be displayed. When `compile` is set
+        to False, the compilation is omitted without any
+        warning.
 
     # Raises
         ImportError: if h5py is not available.
@@ -229,56 +237,58 @@ def load_model(filepath, custom_objects=None):
         if obj in custom_objects:
             return custom_objects[obj]
         return obj
+    with h5py.File(filepath, mode='r') as f:
+        # instantiate model
+        model_config = f.attrs.get('model_config')
+        if model_config is None:
+            raise ValueError('No model found in config file.')
+        model_config = json.loads(model_config.decode('utf-8'))
+        model = model_from_config(model_config, custom_objects=custom_objects)
 
-    f = h5py.File(filepath, mode='r')
+        # set weights
+        topology.load_weights_from_hdf5_group(f['model_weights'], model.layers)
 
-    # instantiate model
-    model_config = f.attrs.get('model_config')
-    if model_config is None:
-        raise ValueError('No model found in config file.')
-    model_config = json.loads(model_config.decode('utf-8'))
-    model = model_from_config(model_config, custom_objects=custom_objects)
+        # Early return if compilation is not required.
+        if not compile:
+            return model
 
-    # set weights
-    topology.load_weights_from_hdf5_group(f['model_weights'], model.layers)
+        # instantiate optimizer
+        training_config = f.attrs.get('training_config')
+        if training_config is None:
+            warnings.warn('No training configuration found in save file: '
+                          'the model was *not* compiled. Compile it manually.')
+            return model
+        training_config = json.loads(training_config.decode('utf-8'))
+        optimizer_config = training_config['optimizer_config']
+        optimizer = optimizers.deserialize(optimizer_config,
+                                           custom_objects=custom_objects)
 
-    # instantiate optimizer
-    training_config = f.attrs.get('training_config')
-    if training_config is None:
-        warnings.warn('No training configuration found in save file: '
-                      'the model was *not* compiled. Compile it manually.')
-        f.close()
-        return model
-    training_config = json.loads(training_config.decode('utf-8'))
-    optimizer_config = training_config['optimizer_config']
-    optimizer = optimizers.deserialize(optimizer_config,
-                                       custom_objects=custom_objects)
+        # Recover loss functions and metrics.
+        loss = convert_custom_objects(training_config['loss'])
+        metrics = convert_custom_objects(training_config['metrics'])
+        sample_weight_mode = training_config['sample_weight_mode']
+        loss_weights = training_config['loss_weights']
 
-    # Recover loss functions and metrics.
-    loss = convert_custom_objects(training_config['loss'])
-    metrics = convert_custom_objects(training_config['metrics'])
-    sample_weight_mode = training_config['sample_weight_mode']
-    loss_weights = training_config['loss_weights']
+        # Compile model.
+        model.compile(optimizer=optimizer,
+                      loss=loss,
+                      metrics=metrics,
+                      loss_weights=loss_weights,
+                      sample_weight_mode=sample_weight_mode)
 
-    # Compile model.
-    model.compile(optimizer=optimizer,
-                  loss=loss,
-                  metrics=metrics,
-                  loss_weights=loss_weights,
-                  sample_weight_mode=sample_weight_mode)
-
-    # Set optimizer weights.
-    if 'optimizer_weights' in f:
-        # Build train function (to get weight updates).
-        if isinstance(model, Sequential):
-            model.model._make_train_function()
-        else:
-            model._make_train_function()
-        optimizer_weights_group = f['optimizer_weights']
-        optimizer_weight_names = [n.decode('utf8') for n in optimizer_weights_group.attrs['weight_names']]
-        optimizer_weight_values = [optimizer_weights_group[n] for n in optimizer_weight_names]
-        model.optimizer.set_weights(optimizer_weight_values)
-    f.close()
+        # Set optimizer weights.
+        if 'optimizer_weights' in f:
+            # Build train function (to get weight updates).
+            if isinstance(model, Sequential):
+                model.model._make_train_function()
+            else:
+                model._make_train_function()
+            optimizer_weights_group = f['optimizer_weights']
+            optimizer_weight_names = [n.decode('utf8') for n in
+                                      optimizer_weights_group.attrs['weight_names']]
+            optimizer_weight_values = [optimizer_weights_group[n] for n in
+                                       optimizer_weight_names]
+            model.optimizer.set_weights(optimizer_weight_values)
     return model
 
 
@@ -293,9 +303,12 @@ def model_from_config(config, custom_objects=None):
 
     # Returns
         A Keras model instance (uncompiled).
+
+    # Raises
+        TypeError: if `config` is not a dictionary.
     """
     if isinstance(config, list):
-        raise TypeError('`model_fom_config` expects a dictionary, not a list. '
+        raise TypeError('`model_from_config` expects a dictionary, not a list. '
                         'Maybe you meant to use '
                         '`Sequential.from_config(config)`?')
     return layer_module.deserialize(config, custom_objects=custom_objects)
@@ -744,7 +757,7 @@ class Sequential(Model):
             optimizer: str (name of optimizer) or optimizer object.
                 See [optimizers](/optimizers).
             loss: str (name of objective function) or objective function.
-                See [objectives](/objectives).
+                See [losses](/losses).
             metrics: list of metrics to be evaluated by the model
                 during training and testing.
                 Typically you will use `metrics=['accuracy']`.
@@ -753,7 +766,8 @@ class Sequential(Model):
                 sample weighting (2D weights), set this to "temporal".
                 "None" defaults to sample-wise weights (1D).
             **kwargs: for Theano backend, these are passed into K.function.
-                Ignored for Tensorflow backend.
+                When using the Tensorflow backend, these are passed into
+                `tf.Session.run`.
 
         # Example
             ```python
@@ -774,11 +788,14 @@ class Sequential(Model):
                            **kwargs)
         self.optimizer = self.model.optimizer
         self.loss = self.model.loss
+        self.total_loss = self.model.total_loss
         self.loss_weights = self.model.loss_weights
         self.metrics = self.model.metrics
         self.metrics_tensors = self.model.metrics_tensors
         self.metrics_names = self.model.metrics_names
         self.sample_weight_mode = self.model.sample_weight_mode
+        self.sample_weights = self.model.sample_weights
+        self.targets = self.model.targets
 
     def fit(self, x, y, batch_size=32, epochs=10, verbose=1, callbacks=None,
             validation_split=0., validation_data=None, shuffle=True,
@@ -1033,8 +1050,8 @@ class Sequential(Model):
                 - a tuple (inputs, targets, sample_weights).
                 All arrays should contain the same number of samples.
                 The generator is expected to loop over its data
-                indefinitely. An epoch finishes when `samples_per_epoch`
-                samples have been seen by the model.
+                indefinitely. An epoch finishes when `steps_per_epoch`
+                batches have been seen by the model.
             steps_per_epoch: Total number of steps (batches of samples)
                 to yield from `generator` before declaring one epoch
                 finished and starting the next epoch. It should typically
@@ -1087,7 +1104,7 @@ class Sequential(Model):
                         f.close()
 
             model.fit_generator(generate_arrays_from_file('/my_file.txt'),
-                                samples_per_epoch=10000, epochs=10)
+                                steps_per_epoch=1000, epochs=10)
         ```
         """
         if self.model is None:
@@ -1227,6 +1244,15 @@ class Sequential(Model):
 
     @classmethod
     def legacy_from_config(cls, config, layer_cache=None):
+        """Load a model from a legacy configuration.
+
+        # Arguments
+            config: dictionary with configuration.
+            layer_cache: cache to draw pre-existing layer.
+
+        # Returns
+            The loaded Model.
+        """
         if not layer_cache:
             layer_cache = {}
 

keras/optimizers.py

@@ -1,5 +1,6 @@
 from __future__ import absolute_import
 import six
+import copy
 from six.moves import zip
 
 from . import backend as K
@@ -11,8 +12,31 @@ if K.backend() == 'tensorflow':
 
 
 def clip_norm(g, c, n):
-    if c > 0:
-        g = K.switch(n >= c, g * c / n, g)
+    if c <= 0:  # if clipnorm == 0 no need to add ops to the graph
+        return g
+
+    # tf require using a special op to multiply IndexedSliced by scalar
+    if K.backend() == 'tensorflow':
+        condition = n >= c
+        then_expression = tf.scalar_mul(c / n, g)
+        else_expression = g
+
+        # saving the shape to avoid converting sparse tensor to dense
+        if isinstance(then_expression, tf.Tensor):
+            g_shape = copy.copy(then_expression.get_shape())
+        elif isinstance(then_expression, tf.IndexedSlices):
+            g_shape = copy.copy(then_expression.dense_shape)
+        if condition.dtype != tf.bool:
+            condition = tf.cast(condition, 'bool')
+        g = tf.cond(condition,
+                    lambda: then_expression,
+                    lambda: else_expression)
+        if isinstance(then_expression, tf.Tensor):
+            g.set_shape(g_shape)
+        elif isinstance(then_expression, tf.IndexedSlices):
+            g._dense_shape = g_shape
+    else:
+        g = K.switch(K.greater_equal(n, c), g * c / n, g)
     return g
 
 

keras/preprocessing/image.py

@@ -13,6 +13,8 @@ from six.moves import range
 import os
 import threading
 import warnings
+import multiprocessing.pool
+from functools import partial
 
 from .. import backend as K
 
@@ -325,9 +327,9 @@ def load_img(path, grayscale=False, target_size=None):
         if img.mode != 'RGB':
             img = img.convert('RGB')
     if target_size:
-        wh_tuple = (target_size[1], target_size[0])
-        if img.size != wh_tuple:
-            img = img.resize(wh_tuple)
+        hw_tuple = (target_size[1], target_size[0])
+        if img.size != hw_tuple:
+            img = img.resize(hw_tuple)
     return img
 
 
@@ -346,6 +348,7 @@ class ImageDataGenerator(object):
         featurewise_std_normalization: divide inputs by std of the dataset.
         samplewise_std_normalization: divide each input by its std.
         zca_whitening: apply ZCA whitening.
+        zca_epsilon: epsilon for ZCA whitening. Default is 1e-6.
         rotation_range: degrees (0 to 180).
         width_shift_range: fraction of total width.
         height_shift_range: fraction of total height.
@@ -382,6 +385,7 @@ class ImageDataGenerator(object):
                  featurewise_std_normalization=False,
                  samplewise_std_normalization=False,
                  zca_whitening=False,
+                 zca_epsilon=1e-6,
                  rotation_range=0.,
                  width_shift_range=0.,
                  height_shift_range=0.,
@@ -402,6 +406,7 @@ class ImageDataGenerator(object):
         self.featurewise_std_normalization = featurewise_std_normalization
         self.samplewise_std_normalization = samplewise_std_normalization
         self.zca_whitening = zca_whitening
+        self.zca_epsilon = zca_epsilon
         self.rotation_range = rotation_range
         self.width_shift_range = width_shift_range
         self.height_shift_range = height_shift_range
@@ -443,7 +448,7 @@ class ImageDataGenerator(object):
                              'Received arg: ', zoom_range)
 
     def flow(self, x, y=None, batch_size=32, shuffle=True, seed=None,
-             save_to_dir=None, save_prefix='', save_format='jpeg'):
+             save_to_dir=None, save_prefix='', save_format='png'):
         return NumpyArrayIterator(
             x, y, self,
             batch_size=batch_size,
@@ -460,7 +465,7 @@ class ImageDataGenerator(object):
                             batch_size=32, shuffle=True, seed=None,
                             save_to_dir=None,
                             save_prefix='',
-                            save_format='jpeg',
+                            save_format='png',
                             follow_links=False):
         return DirectoryIterator(
             directory, self,
@@ -633,8 +638,8 @@ class ImageDataGenerator(object):
         if x.ndim != 4:
             raise ValueError('Input to `.fit()` should have rank 4. '
                              'Got array with shape: ' + str(x.shape))
-        if x.shape[self.channel_axis] not in {1, 3, 4}:
-            raise ValueError(
+        if x.shape[self.channel_axis] not in {3, 4}:
+            warnings.warn(
                 'Expected input to be images (as Numpy array) '
                 'following the data format convention "' + self.data_format + '" '
                 '(channels on axis ' + str(self.channel_axis) + '), i.e. expected '
@@ -671,7 +676,7 @@ class ImageDataGenerator(object):
             flat_x = np.reshape(x, (x.shape[0], x.shape[1] * x.shape[2] * x.shape[3]))
             sigma = np.dot(flat_x.T, flat_x) / flat_x.shape[0]
             u, s, _ = linalg.svd(sigma)
-            self.principal_components = np.dot(np.dot(u, np.diag(1. / np.sqrt(s + 10e-7))), u.T)
+            self.principal_components = np.dot(np.dot(u, np.diag(1. / np.sqrt(s + self.zca_epsilon))), u.T)
 
 
 class Iterator(object):
@@ -752,7 +757,7 @@ class NumpyArrayIterator(Iterator):
     def __init__(self, x, y, image_data_generator,
                  batch_size=32, shuffle=False, seed=None,
                  data_format=None,
-                 save_to_dir=None, save_prefix='', save_format='jpeg'):
+                 save_to_dir=None, save_prefix='', save_format='png'):
         if y is not None and len(x) != len(y):
             raise ValueError('X (images tensor) and y (labels) '
                              'should have the same length. '
@@ -818,6 +823,73 @@ class NumpyArrayIterator(Iterator):
         return batch_x, batch_y
 
 
+def _count_valid_files_in_directory(directory, white_list_formats, follow_links):
+    """Count files with extension in `white_list_formats` contained in a directory.
+
+    # Arguments
+        directory: absolute path to the directory containing files to be counted
+        white_list_formats: set of strings containing allowed extensions for
+            the files to be counted.
+
+    # Returns
+        the count of files with extension in `white_list_formats` contained in
+        the directory.
+    """
+    def _recursive_list(subpath):
+        return sorted(os.walk(subpath, followlinks=follow_links), key=lambda tpl: tpl[0])
+
+    samples = 0
+    for root, _, files in _recursive_list(directory):
+        for fname in files:
+            is_valid = False
+            for extension in white_list_formats:
+                if fname.lower().endswith('.' + extension):
+                    is_valid = True
+                    break
+            if is_valid:
+                samples += 1
+    return samples
+
+
+def _list_valid_filenames_in_directory(directory, white_list_formats,
+                                       class_indices, follow_links):
+    """List paths of files in `subdir` relative from `directory` whose extensions are in `white_list_formats`.
+
+    # Arguments
+        directory: absolute path to a directory containing the files to list.
+            The directory name is used as class label and must be a key of `class_indices`.
+        white_list_formats: set of strings containing allowed extensions for
+            the files to be counted.
+        class_indices: dictionary mapping a class name to its index.
+
+    # Returns
+        classes: a list of class indices
+        filenames: the path of valid files in `directory`, relative from
+            `directory`'s parent (e.g., if `directory` is "dataset/class1",
+            the filenames will be ["class1/file1.jpg", "class1/file2.jpg", ...]).
+    """
+    def _recursive_list(subpath):
+        return sorted(os.walk(subpath, followlinks=follow_links), key=lambda tpl: tpl[0])
+
+    classes = []
+    filenames = []
+    subdir = os.path.basename(directory)
+    basedir = os.path.dirname(directory)
+    for root, _, files in _recursive_list(directory):
+        for fname in files:
+            is_valid = False
+            for extension in white_list_formats:
+                if fname.lower().endswith('.' + extension):
+                    is_valid = True
+                    break
+            if is_valid:
+                classes.append(class_indices[subdir])
+                # add filename relative to directory
+                absolute_path = os.path.join(root, fname)
+                filenames.append(os.path.relpath(absolute_path, basedir))
+    return classes, filenames
+
+
 class DirectoryIterator(Iterator):
     """Iterator capable of reading images from a directory on disk.
 
@@ -838,6 +910,8 @@ class DirectoryIterator(Iterator):
             `"binary"`: binary targets (if there are only two classes),
             `"categorical"`: categorical targets,
             `"sparse"`: integer targets,
+            `"input"`: targets are images identical to input images (mainly
+                used to work with autoencoders),
             `None`: no targets get yielded (only input images are yielded).
         batch_size: Integer, size of a batch.
         shuffle: Boolean, whether to shuffle the data between epochs.
@@ -858,7 +932,7 @@ class DirectoryIterator(Iterator):
                  classes=None, class_mode='categorical',
                  batch_size=32, shuffle=True, seed=None,
                  data_format=None,
-                 save_to_dir=None, save_prefix='', save_format='jpeg',
+                 save_to_dir=None, save_prefix='', save_format='png',
                  follow_links=False):
         if data_format is None:
             data_format = K.image_data_format()
@@ -881,10 +955,12 @@ class DirectoryIterator(Iterator):
             else:
                 self.image_shape = (1,) + self.target_size
         self.classes = classes
-        if class_mode not in {'categorical', 'binary', 'sparse', None}:
+        if class_mode not in {'categorical', 'binary', 'sparse',
+                              'input', None}:
             raise ValueError('Invalid class_mode:', class_mode,
                              '; expected one of "categorical", '
-                             '"binary", "sparse", or None.')
+                             '"binary", "sparse", "input"'
+                             ' or None.')
         self.class_mode = class_mode
         self.save_to_dir = save_to_dir
         self.save_prefix = save_prefix
@@ -906,38 +982,33 @@ class DirectoryIterator(Iterator):
         def _recursive_list(subpath):
             return sorted(os.walk(subpath, followlinks=follow_links), key=lambda tpl: tpl[0])
 
-        for subdir in classes:
-            subpath = os.path.join(directory, subdir)
-            for root, _, files in _recursive_list(subpath):
-                for fname in files:
-                    is_valid = False
-                    for extension in white_list_formats:
-                        if fname.lower().endswith('.' + extension):
-                            is_valid = True
-                            break
-                    if is_valid:
-                        self.samples += 1
+        pool = multiprocessing.pool.ThreadPool()
+        function_partial = partial(_count_valid_files_in_directory,
+                                   white_list_formats=white_list_formats,
+                                   follow_links=follow_links)
+        self.samples = sum(pool.map(function_partial,
+                                    (os.path.join(directory, subdir)
+                                     for subdir in classes)))
+
         print('Found %d images belonging to %d classes.' % (self.samples, self.num_class))
 
         # second, build an index of the images in the different class subfolders
+        results = []
+
         self.filenames = []
         self.classes = np.zeros((self.samples,), dtype='int32')
         i = 0
-        for subdir in classes:
-            subpath = os.path.join(directory, subdir)
-            for root, _, files in _recursive_list(subpath):
-                for fname in files:
-                    is_valid = False
-                    for extension in white_list_formats:
-                        if fname.lower().endswith('.' + extension):
-                            is_valid = True
-                            break
-                    if is_valid:
-                        self.classes[i] = self.class_indices[subdir]
-                        i += 1
-                        # add filename relative to directory
-                        absolute_path = os.path.join(root, fname)
-                        self.filenames.append(os.path.relpath(absolute_path, directory))
+        for dirpath in (os.path.join(directory, subdir) for subdir in classes):
+            results.append(pool.apply_async(_list_valid_filenames_in_directory,
+                                            (dirpath, white_list_formats,
+                                             self.class_indices, follow_links)))
+        for res in results:
+            classes, filenames = res.get()
+            self.classes[i:i + len(classes)] = classes
+            self.filenames += filenames
+            i += len(classes)
+        pool.close()
+        pool.join()
         super(DirectoryIterator, self).__init__(self.samples, batch_size, shuffle, seed)
 
     def next(self):
@@ -972,7 +1043,9 @@ class DirectoryIterator(Iterator):
                                                                   format=self.save_format)
                 img.save(os.path.join(self.save_to_dir, fname))
         # build batch of labels
-        if self.class_mode == 'sparse':
+        if self.class_mode == 'input':
+            batch_y = batch_x.copy()
+        elif self.class_mode == 'sparse':
             batch_y = self.classes[index_array]
         elif self.class_mode == 'binary':
             batch_y = self.classes[index_array].astype(K.floatx())

keras/preprocessing/text.py

@@ -11,6 +11,7 @@ import sys
 import numpy as np
 from six.moves import range
 from six.moves import zip
+from collections import OrderedDict
 import warnings
 
 if sys.version_info < (3,):
@@ -22,7 +23,7 @@ else:
 def text_to_word_sequence(text,
                           filters='!"#$%&()*+,-./:;<=>?@[\\]^_`{|}~\t\n',
                           lower=True, split=" "):
-    """Converts a text to a sequence of word indices.
+    """Converts a text to a sequence of words (or tokens).
 
     # Arguments
         text: Input text (string).
@@ -31,7 +32,7 @@ def text_to_word_sequence(text,
         split: Sentence split marker (string).
 
     # Returns
-        A list of integer word indices.
+        A list of words (or tokens).
     """
     if lower:
         text = text.lower()
@@ -68,7 +69,7 @@ class Tokenizer(object):
             tabs and line breaks, minus the `'` character.
         lower: boolean. Whether to convert the texts to lowercase.
         split: character or string to use for token splitting.
-        char_level: if True, every character will be treated as a word.
+        char_level: if True, every character will be treated as a token.
 
     By default, all punctuation is removed, turning the texts into
     space-separated sequences of words
@@ -92,7 +93,7 @@ class Tokenizer(object):
         if kwargs:
             raise TypeError('Unrecognized keyword arguments: ' + str(kwargs))
 
-        self.word_counts = {}
+        self.word_counts = OrderedDict()
         self.word_docs = {}
         self.filters = filters
         self.split = split

keras/utils/conv_utils.py

@@ -70,7 +70,7 @@ def convert_kernel(kernel):
     Also works reciprocally, since the transformation is its own inverse.
 
     # Arguments
-        kernel: Numpy array (4D or 5D).
+        kernel: Numpy array (3D, 4D or 5D).
 
     # Returns
         The converted kernel.
@@ -78,7 +78,8 @@ def convert_kernel(kernel):
     # Raises
         ValueError: in case of invalid kernel shape or invalid data_format.
     """
-    if not 4 <= kernel.ndim <= 5:
+    kernel = np.asarray(kernel)
+    if not 3 <= kernel.ndim <= 5:
         raise ValueError('Invalid kernel shape:', kernel.shape)
     slices = [slice(None, None, -1) for _ in range(kernel.ndim)]
     no_flip = (slice(None, None), slice(None, None))

keras/utils/data_utils.py

@@ -2,7 +2,6 @@
 from __future__ import absolute_import
 from __future__ import print_function
 
-import functools
 import tarfile
 import zipfile
 import os
@@ -36,8 +35,10 @@ if sys.version_info[0] == 2:
             data: `data` argument passed to `urlopen`.
         """
         def chunk_read(response, chunk_size=8192, reporthook=None):
-            total_size = response.info().get('Content-Length').strip()
-            total_size = int(total_size)
+            content_type = response.info().get('Content-Length')
+            total_size = -1
+            if content_type is not None:
+                total_size = int(content_type.strip())
             count = 0
             while 1:
                 chunk = response.read(chunk_size)
@@ -186,19 +187,24 @@ def get_file(fname,
 
     if download:
         print('Downloading data from', origin)
-        progbar = None
 
-        def dl_progress(count, block_size, total_size, progbar=None):
-            if progbar is None:
-                progbar = Progbar(total_size)
+        class ProgressTracker(object):
+            # Maintain progbar for the lifetime of download.
+            # This design was chosen for Python 2.7 compatibility.
+            progbar = None
+
+        def dl_progress(count, block_size, total_size):
+            if ProgressTracker.progbar is None:
+                if total_size is -1:
+                    total_size = None
+                ProgressTracker.progbar = Progbar(total_size)
             else:
-                progbar.update(count * block_size)
+                ProgressTracker.progbar.update(count * block_size)
 
         error_msg = 'URL fetch failure on {}: {} -- {}'
         try:
             try:
-                urlretrieve(origin, fpath,
-                            functools.partial(dl_progress, progbar=progbar))
+                urlretrieve(origin, fpath, dl_progress)
             except URLError as e:
                 raise Exception(error_msg.format(origin, e.errno, e.reason))
             except HTTPError as e:
@@ -207,7 +213,7 @@ def get_file(fname,
             if os.path.exists(fpath):
                 os.remove(fpath)
             raise
-        progbar = None
+        ProgressTracker.progbar = None
 
     if untar:
         if not os.path.exists(untar_fpath):

keras/utils/generic_utils.py

@@ -27,8 +27,8 @@ class CustomObjectScope(object):
     Consider a custom object `MyObject`
 
     ```python
-        with CustomObjectScope({"MyObject":MyObject}):
-            layer = Dense(..., W_regularizer="MyObject")
+        with CustomObjectScope({'MyObject':MyObject}):
+            layer = Dense(..., kernel_regularizer='MyObject')
             # save, load, etc. will recognize custom object by name
     ```
     """
@@ -63,8 +63,8 @@ def custom_object_scope(*args):
     Consider a custom object `MyObject`
 
     ```python
-        with custom_object_scope({"MyObject":MyObject}):
-            layer = Dense(..., W_regularizer="MyObject")
+        with custom_object_scope({'MyObject':MyObject}):
+            layer = Dense(..., kernel_regularizer='MyObject')
             # save, load, etc. will recognize custom object by name
     ```
 
@@ -89,7 +89,7 @@ def get_custom_objects():
 
     ```python
         get_custom_objects().clear()
-        get_custom_objects()["MyObject"] = MyObject
+        get_custom_objects()['MyObject'] = MyObject
     ```
 
     # Returns
@@ -133,17 +133,20 @@ def deserialize_keras_object(identifier, module_objects=None,
                                  ': ' + class_name)
         if hasattr(cls, 'from_config'):
             arg_spec = inspect.getargspec(cls.from_config)
+            custom_objects = custom_objects or {}
             if 'custom_objects' in arg_spec.args:
-                custom_objects = custom_objects or {}
                 return cls.from_config(config['config'],
                                        custom_objects=dict(list(_GLOBAL_CUSTOM_OBJECTS.items()) +
                                                            list(custom_objects.items())))
-            return cls.from_config(config['config'])
+            with CustomObjectScope(custom_objects):
+                return cls.from_config(config['config'])
         else:
             # Then `cls` may be a function returning a class.
             # in this case by convention `config` holds
             # the kwargs of the function.
-            return cls(**config['config'])
+            custom_objects = custom_objects or {}
+            with CustomObjectScope(custom_objects):
+                return cls(**config['config'])
     elif isinstance(identifier, six.string_types):
         function_name = identifier
         if custom_objects and function_name in custom_objects:
@@ -153,7 +156,7 @@ def deserialize_keras_object(identifier, module_objects=None,
         else:
             fn = module_objects.get(function_name)
             if fn is None:
-                raise ValueError('Unknown ' + printable_module_name,
+                raise ValueError('Unknown ' + printable_module_name +
                                  ':' + function_name)
         return fn
     else:
@@ -208,12 +211,14 @@ class Progbar(object):
     """Displays a progress bar.
 
     # Arguments
-        target: Total number of steps expected.
+        target: Total number of steps expected, None if unknown.
         interval: Minimum visual progress update interval (in seconds).
     """
 
     def __init__(self, target, width=30, verbose=1, interval=0.05):
         self.width = width
+        if target is None:
+            target = -1
         self.target = target
         self.sum_values = {}
         self.unique_values = []
@@ -253,21 +258,22 @@ class Progbar(object):
             sys.stdout.write('\b' * prev_total_width)
             sys.stdout.write('\r')
 
-            numdigits = int(np.floor(np.log10(self.target))) + 1
-            barstr = '%%%dd/%%%dd [' % (numdigits, numdigits)
-            bar = barstr % (current, self.target)
-            prog = float(current) / self.target
-            prog_width = int(self.width * prog)
-            if prog_width > 0:
-                bar += ('=' * (prog_width - 1))
-                if current < self.target:
-                    bar += '>'
-                else:
-                    bar += '='
-            bar += ('.' * (self.width - prog_width))
-            bar += ']'
-            sys.stdout.write(bar)
-            self.total_width = len(bar)
+            if self.target is not -1:
+                numdigits = int(np.floor(np.log10(self.target))) + 1
+                barstr = '%%%dd/%%%dd [' % (numdigits, numdigits)
+                bar = barstr % (current, self.target)
+                prog = float(current) / self.target
+                prog_width = int(self.width * prog)
+                if prog_width > 0:
+                    bar += ('=' * (prog_width - 1))
+                    if current < self.target:
+                        bar += '>'
+                    else:
+                        bar += '='
+                bar += ('.' * (self.width - prog_width))
+                bar += ']'
+                sys.stdout.write(bar)
+                self.total_width = len(bar)
 
             if current:
                 time_per_unit = (now - self.start) / current
@@ -275,7 +281,7 @@ class Progbar(object):
                 time_per_unit = 0
             eta = time_per_unit * (self.target - current)
             info = ''
-            if current < self.target:
+            if current < self.target and self.target is not -1:
                 info += ' - ETA: %ds' % eta
             else:
                 info += ' - %ds' % (now - self.start)

keras/utils/io_utils.py

@@ -63,8 +63,13 @@ class HDF5Matrix(object):
 
     def __getitem__(self, key):
         if isinstance(key, slice):
-            if key.stop + self.start <= self.end:
-                idx = slice(key.start + self.start, key.stop + self.start)
+            start, stop = key.start, key.stop
+            if start is None:
+                start = 0
+            if stop is None:
+                stop = self.data.shape[0]
+            if stop + self.start <= self.end:
+                idx = slice(start + self.start, stop + self.start)
             else:
                 raise IndexError
         elif isinstance(key, int):

keras/utils/layer_utils.py

@@ -19,8 +19,11 @@ def print_summary(model, line_length=None, positions=None):
     else:
         sequential_like = True
         for v in model.nodes_by_depth.values():
-            if len(v) > 1:
+            if (len(v) > 1) or (len(v) == 1 and len(v[0].inbound_layers) > 1):
+                # if the model has multiple nodes or if the nodes have multiple inbound_layers
+                # the model is no longer sequential
                 sequential_like = False
+                break
 
     if sequential_like:
         line_length = line_length or 65
@@ -75,12 +78,10 @@ def print_summary(model, line_length=None, positions=None):
         except AttributeError:
             output_shape = 'multiple'
         connections = []
-        for node_index, node in enumerate(layer.inbound_nodes):
-            if relevant_nodes:
-                node_key = layer.name + '_ib-' + str(node_index)
-                if node_key not in relevant_nodes:
-                    # node is node part of the current network
-                    continue
+        for node in layer.inbound_nodes:
+            if relevant_nodes and node not in relevant_nodes:
+                # node is not part of the current network
+                continue
             for i in range(len(node.inbound_layers)):
                 inbound_layer = node.inbound_layers[i].name
                 inbound_node_index = node.node_indices[i]
@@ -111,7 +112,10 @@ def print_summary(model, line_length=None, positions=None):
         else:
             print('_' * line_length)
 
-    trainable_count, non_trainable_count = count_total_params(layers, layer_set=None)
+    trainable_count = int(
+        np.sum([K.count_params(p) for p in set(model.trainable_weights)]))
+    non_trainable_count = int(
+        np.sum([K.count_params(p) for p in set(model.non_trainable_weights)]))
 
     print('Total params: {:,}'.format(trainable_count + non_trainable_count))
     print('Trainable params: {:,}'.format(trainable_count))
@@ -119,35 +123,6 @@ def print_summary(model, line_length=None, positions=None):
     print('_' * line_length)
 
 
-def count_total_params(layers, layer_set=None):
-    """Counts the number of parameters in a list of layers.
-
-    # Arguments
-        layers: list of layers.
-        layer_set: set of layers already seen
-            (so that we don't count their weights twice).
-
-    # Returns
-        A tuple (count of trainable weights, count of non-trainable weights.)
-    """
-    if layer_set is None:
-        layer_set = set()
-    trainable_count = 0
-    non_trainable_count = 0
-    for layer in layers:
-        if layer in layer_set:
-            continue
-        layer_set.add(layer)
-        if hasattr(layer, 'layers'):
-            t, nt = count_total_params(layer.layers, layer_set)
-            trainable_count += t
-            non_trainable_count += nt
-        else:
-            trainable_count += np.sum([K.count_params(p) for p in layer.trainable_weights])
-            non_trainable_count += np.sum([K.count_params(p) for p in layer.non_trainable_weights])
-    return int(trainable_count), int(non_trainable_count)
-
-
 def convert_all_kernels_in_model(model):
     """Converts all convolution kernels in a model from Theano to TensorFlow.
 
@@ -192,7 +167,7 @@ def convert_dense_weights_data_format(dense,
             came before the target `Dense` layer.
         target_data_format: One of "channels_last", "channels_first".
             Set it "channels_last"
-            if converting a "chnnels_first" model to "channels_last",
+            if converting a "channels_first" model to "channels_last",
             or reciprocally.
     """
     assert target_data_format in {'channels_last', 'channels_first'}

keras/utils/vis_utils.py

@@ -5,26 +5,43 @@ try:
     # pydot-ng is a fork of pydot that is better maintained.
     import pydot_ng as pydot
 except ImportError:
-    # Fall back on pydot if necessary.
+    # pydotplus is an improved version of pydot
     try:
-        import pydot
+        import pydotplus as pydot
     except ImportError:
-        pydot = None
+        # Fall back on pydot if necessary.
+        try:
+            import pydot
+        except ImportError:
+            pydot = None
 
 
 def _check_pydot():
-    if not (pydot and pydot.find_graphviz()):
+    try:
+        # Attempt to create an image of a blank graph
+        # to check the pydot/graphviz installation.
+        pydot.Dot.create(pydot.Dot())
+    except Exception:
+        # pydot raises a generic Exception here,
+        # so no specific class can be caught.
         raise ImportError('Failed to import pydot. You must install pydot'
                           ' and graphviz for `pydotprint` to work.')
 
 
-def model_to_dot(model, show_shapes=False, show_layer_names=True):
-    """Converts a Keras model to dot format.
+def model_to_dot(model,
+                 show_shapes=False,
+                 show_layer_names=True,
+                 rankdir='TB'):
+    """Convert a Keras model to dot format.
 
     # Arguments
         model: A Keras model instance.
         show_shapes: whether to display shape information.
         show_layer_names: whether to display layer names.
+        rankdir: `rankdir` argument passed to PyDot,
+            a string specifying the format of the plot:
+            'TB' creates a vertical plot;
+            'LR' creates a horizontal plot.
 
     # Returns
         A `pydot.Dot` instance representing the Keras model.
@@ -34,7 +51,7 @@ def model_to_dot(model, show_shapes=False, show_layer_names=True):
 
     _check_pydot()
     dot = pydot.Dot()
-    dot.set('rankdir', 'TB')
+    dot.set('rankdir', rankdir)
     dot.set('concentrate', True)
     dot.set_node_defaults(shape='record')
 
@@ -75,8 +92,9 @@ def model_to_dot(model, show_shapes=False, show_layer_names=True):
                     [str(ishape) for ishape in layer.input_shapes])
             else:
                 inputlabels = 'multiple'
-            label = '%s\n|{input:|output:}|{{%s}|{%s}}' % (label, inputlabels, outputlabels)
-
+            label = '%s\n|{input:|output:}|{{%s}|{%s}}' % (label,
+                                                           inputlabels,
+                                                           outputlabels)
         node = pydot.Node(layer_id, label=label)
         dot.add_node(node)
 
@@ -96,8 +114,21 @@ def model_to_dot(model, show_shapes=False, show_layer_names=True):
 def plot_model(model,
                to_file='model.png',
                show_shapes=False,
-               show_layer_names=True):
-    dot = model_to_dot(model, show_shapes, show_layer_names)
+               show_layer_names=True,
+               rankdir='TB'):
+    """Converts a Keras model to dot format and save to a file.
+
+    # Arguments
+        model: A Keras model instance
+        to_file: File name of the plot image.
+        show_shapes: whether to display shape information.
+        show_layer_names: whether to display layer names.
+        rankdir: `rankdir` argument passed to PyDot,
+            a string specifying the format of the plot:
+            'TB' creates a vertical plot;
+            'LR' creates a horizontal plot.
+    """
+    dot = model_to_dot(model, show_shapes, show_layer_names, rankdir)
     _, extension = os.path.splitext(to_file)
     if not extension:
         extension = 'png'

keras/wrappers/scikit_learn.py

@@ -90,7 +90,7 @@ class BaseWrapper(object):
         """Gets parameters for this estimator.
 
         # Arguments
-            **params: ignored (exists for API compatiblity).
+            **params: ignored (exists for API compatibility).
 
         # Returns
             Dictionary of parameter names mapped to their values.

setup.py

@@ -3,12 +3,12 @@ from setuptools import find_packages
 
 
 setup(name='Keras',
-      version='2.0.3',
+      version='2.0.5',
       description='Deep Learning for Python',
       author='Francois Chollet',
       author_email='francois.chollet@gmail.com',
       url='https://github.com/fchollet/keras',
-      download_url='https://github.com/fchollet/keras/tarball/2.0.3',
+      download_url='https://github.com/fchollet/keras/tarball/2.0.5',
       license='MIT',
       install_requires=['theano', 'pyyaml', 'six'],
       extras_require={

tests/integration_tests/test_temporal_data_tasks.py

@@ -6,7 +6,8 @@ import string
 from keras.utils.test_utils import get_test_data, keras_test
 from keras.utils.np_utils import to_categorical
 from keras.models import Sequential
-from keras import layers
+from keras import layers, optimizers
+import keras.backend as K
 import keras
 
 
@@ -204,5 +205,14 @@ def test_masked_temporal():
     ground_truth = -np.log(0.5)
     assert(np.abs(history.history['loss'][-1] - ground_truth) < 0.06)
 
+
+@pytest.mark.skipif(K.backend() != 'tensorflow', reason='Requires TF backend')
+@keras_test
+def test_embedding_with_clipnorm():
+    model = Sequential()
+    model.add(layers.Embedding(input_dim=1, output_dim=1))
+    model.compile(optimizer=optimizers.SGD(clipnorm=0.1), loss='mse')
+    model.fit(np.array([[0]]), np.array([[[0.5]]]), epochs=1)
+
 if __name__ == '__main__':
     pytest.main([__file__])

tests/keras/activations_test.py

@@ -136,6 +136,10 @@ def test_elu():
     assert_allclose(result, test_values, rtol=1e-05)
 
     negative_values = np.array([[-1, -2]], dtype=K.floatx())
+    # cntk can't rebind the input shape, so create the model again to test different batch size
+    if (K.backend() == 'cntk'):
+        x2 = K.placeholder(ndim=2)
+        f = K.function([x2], [activations.elu(x2, 0.5)])
     result = f([negative_values])[0]
     true_result = (np.exp(negative_values) - 1) / 2
 

tests/keras/applications/applications_test.py

@@ -11,12 +11,24 @@ def test_resnet50():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_resnet50_notop():
     model = applications.ResNet50(weights=None, include_top=False)
     assert model.output_shape == (None, None, None, 2048)
 
 
 @keras_test
+def test_resnet50_notop_specified_input_shape():
+    input_shape = (3, 300, 300) if K.image_data_format() == 'channels_first' else (300, 300, 3)
+    model = applications.ResNet50(weights=None, include_top=False, input_shape=input_shape)
+    output_shape = (None, 2048, 1, 1) if K.image_data_format() == 'channels_first' else (None, 1, 1, 2048)
+    assert model.output_shape == output_shape
+
+
+@keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_resnet50_pooling():
     model = applications.ResNet50(weights=None,
                                   include_top=False,
@@ -24,6 +36,16 @@ def test_resnet50_pooling():
     assert model.output_shape == (None, 2048)
 
 
+@keras_test
+def test_resnet50_pooling_specified_input_shape():
+    input_shape = (3, 300, 300) if K.image_data_format() == 'channels_first' else (300, 300, 3)
+    model = applications.ResNet50(weights=None,
+                                  include_top=False,
+                                  pooling='avg',
+                                  input_shape=input_shape)
+    assert model.output_shape == (None, 2048)
+
+
 @keras_test
 def test_vgg16():
     model = applications.VGG16(weights=None)
@@ -31,17 +53,36 @@ def test_vgg16():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_vgg16_notop():
     model = applications.VGG16(weights=None, include_top=False)
     assert model.output_shape == (None, None, None, 512)
 
 
 @keras_test
+def test_vgg16_notop_specified_input_shape():
+    input_shape = (3, 300, 300) if K.image_data_format() == 'channels_first' else (300, 300, 3)
+    model = applications.VGG16(weights=None, include_top=False, input_shape=input_shape)
+    output_shape = (None, 512, 9, 9) if K.image_data_format() == 'channels_first' else (None, 9, 9, 512)
+    assert model.output_shape == output_shape
+
+
+@keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_vgg16_pooling():
     model = applications.VGG16(weights=None, include_top=False, pooling='avg')
     assert model.output_shape == (None, 512)
 
 
+@keras_test
+def test_vgg16_pooling_specified_input_shape():
+    input_shape = (3, 300, 300) if K.image_data_format() == 'channels_first' else (300, 300, 3)
+    model = applications.VGG16(weights=None, include_top=False, pooling='avg', input_shape=input_shape)
+    assert model.output_shape == (None, 512)
+
+
 @keras_test
 def test_vgg19():
     model = applications.VGG19(weights=None)
@@ -49,17 +90,36 @@ def test_vgg19():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_vgg19_notop():
-    model = applications.VGG16(weights=None, include_top=False)
+    model = applications.VGG19(weights=None, include_top=False)
     assert model.output_shape == (None, None, None, 512)
 
 
 @keras_test
+def test_vgg19_notop_specified_input_shape():
+    input_shape = (3, 300, 300) if K.image_data_format() == 'channels_first' else (300, 300, 3)
+    model = applications.VGG19(weights=None, include_top=False, input_shape=input_shape)
+    output_shape = (None, 512, 9, 9) if K.image_data_format() == 'channels_first' else (None, 9, 9, 512)
+    assert model.output_shape == output_shape
+
+
+@keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_vgg19_pooling():
     model = applications.VGG16(weights=None, include_top=False, pooling='avg')
     assert model.output_shape == (None, 512)
 
 
+@keras_test
+def test_vgg19_pooling_specified_input_shape():
+    input_shape = (3, 300, 300) if K.image_data_format() == 'channels_first' else (300, 300, 3)
+    model = applications.VGG16(weights=None, include_top=False, pooling='avg', input_shape=input_shape)
+    assert model.output_shape == (None, 512)
+
+
 @keras_test
 @pytest.mark.skipif((K.backend() != 'tensorflow'),
                     reason='Requires tensorflow backend')
@@ -91,12 +151,16 @@ def test_inceptionv3():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_inceptionv3_notop():
     model = applications.InceptionV3(weights=None, include_top=False)
     assert model.output_shape == (None, None, None, 2048)
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support padding with non-concrete dimension")
 def test_inceptionv3_pooling():
     model = applications.InceptionV3(weights=None, include_top=False, pooling='avg')
     assert model.output_shape == (None, 2048)

tests/keras/engine/test_topology.py

@@ -77,6 +77,8 @@ def test_trainable_weights():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support add learning_phase() as input")
 def test_learning_phase():
     a = Input(shape=(32,), name='input_a')
     b = Input(shape=(32,), name='input_b')
@@ -439,12 +441,11 @@ def test_recursion():
     with pytest.raises(Exception) as e:
         Model([j], [m, n])
 
-    # redudant outputs
+    # redundant outputs
     j = Input(shape=(32,), name='input_j')
     k = Input(shape=(32,), name='input_k')
     m, n = model([j, k])
-    # this should work lol
-    # TODO: raise a warning
+    # this should work with a warning
     Model([j, k], [m, n, n])
 
     # redundant inputs
@@ -489,6 +490,68 @@ def test_recursion():
         Dense(2)(x)
 
 
+@keras_test
+def test_load_layers():
+    from keras.layers import ConvLSTM2D, TimeDistributed, Bidirectional, Conv2D, Input
+    from keras.models import Model
+    from keras.engine.topology import preprocess_weights_for_loading
+
+    if K.backend() == 'tensorflow' or K.backend() == 'cntk':
+        inputs = Input(shape=(10, 20, 20, 1))
+    else:
+        inputs = Input(shape=(10, 1, 20, 20))
+    td_conv = TimeDistributed(Conv2D(15, (5, 5)))(inputs)
+    bi_convlstm2d = Bidirectional(ConvLSTM2D(10, (3, 3)), merge_mode='concat')(td_conv)
+    model = Model(inputs=inputs, outputs=bi_convlstm2d)
+
+    weight_value_tuples = []
+
+    # TimeDistributed Conv2D layer
+    # use 'channels_first' data format to check that the function is being called correctly for Conv2D
+    # old: (filters, stack_size, kernel_rows, kernel_cols)
+    # new: (kernel_rows, kernel_cols, stack_size, filters)
+    weight_tensor_td_conv_old = list()
+    weight_tensor_td_conv_old.append(np.zeros((15, 1, 5, 5)))
+    weight_tensor_td_conv_old.append(np.zeros((15,)))
+    td_conv_layer = model.layers[1]
+    td_conv_layer.layer.data_format = 'channels_first'
+    weight_tensor_td_conv_new = preprocess_weights_for_loading(td_conv_layer,
+                                                               weight_tensor_td_conv_old,
+                                                               original_keras_version='1')
+    symbolic_weights = td_conv_layer.weights
+    assert (len(symbolic_weights) == len(weight_tensor_td_conv_new))
+    weight_value_tuples += zip(symbolic_weights, weight_tensor_td_conv_new)
+
+    # Bidirectional ConvLSTM2D layer
+    # old ConvLSTM2D took a list of 12 weight tensors, returns a list of 3 concatenated larger tensors.
+    weight_tensor_bi_convlstm_old = []
+    for j in range(2):  # bidirectional
+        for i in range(4):
+            weight_tensor_bi_convlstm_old.append(np.zeros((3, 3, 15, 10)))  # kernel
+            weight_tensor_bi_convlstm_old.append(np.zeros((3, 3, 10, 10)))  # recurrent kernel
+            weight_tensor_bi_convlstm_old.append(np.zeros((10,)))  # bias
+
+    bi_convlstm_layer = model.layers[2]
+    weight_tensor_bi_convlstm_new = preprocess_weights_for_loading(bi_convlstm_layer,
+                                                                   weight_tensor_bi_convlstm_old,
+                                                                   original_keras_version='1')
+
+    symbolic_weights = bi_convlstm_layer.weights
+    assert (len(symbolic_weights) == len(weight_tensor_bi_convlstm_new))
+    weight_value_tuples += zip(symbolic_weights, weight_tensor_bi_convlstm_new)
+
+    K.batch_set_value(weight_value_tuples)
+
+    assert np.all(K.eval(model.layers[1].weights[0]) == weight_tensor_td_conv_new[0])
+    assert np.all(K.eval(model.layers[1].weights[1]) == weight_tensor_td_conv_new[1])
+    assert np.all(K.eval(model.layers[2].weights[0]) == weight_tensor_bi_convlstm_new[0])
+    assert np.all(K.eval(model.layers[2].weights[1]) == weight_tensor_bi_convlstm_new[1])
+    assert np.all(K.eval(model.layers[2].weights[2]) == weight_tensor_bi_convlstm_new[2])
+    assert np.all(K.eval(model.layers[2].weights[3]) == weight_tensor_bi_convlstm_new[3])
+    assert np.all(K.eval(model.layers[2].weights[4]) == weight_tensor_bi_convlstm_new[4])
+    assert np.all(K.eval(model.layers[2].weights[5]) == weight_tensor_bi_convlstm_new[5])
+
+
 @keras_test
 def test_recursion_with_bn_and_loss():
     model1 = Sequential([
@@ -516,5 +579,32 @@ def test_recursion_with_bn_and_loss():
     model2.fit(x, y, verbose=0, epochs=1)
 
 
+def test_shared_layer_depth_is_correct():
+    # Basic outline here: we have a shared embedding layer, and two inputs that go through
+    # different depths of computation in the graph before the final output.  We need the computed
+    # depth of the input layers to be the same, because they both pass through the embedding layer
+    # before anything else happens.  That's what we're testing.
+    from keras.layers import Embedding, Input, Dense, Concatenate
+    from keras.models import Model
+    input1 = Input(shape=(10,), name="input1")
+    input2 = Input(shape=(10,), name="input2")
+    embedding_layer = Embedding(name="embedding", input_dim=5, output_dim=10)
+    embedded_input1 = embedding_layer(input1)
+    embedded_input2 = embedding_layer(input2)
+    transformed_input2 = Dense(6)(Dense(5)(Dense(3)(embedded_input2)))
+    final_output = Dense(2)(Concatenate()([embedded_input1, transformed_input2]))
+    model = Model(inputs=[input1, input2], outputs=final_output)
+    input1_depth = -1
+    input2_depth = -1
+    for depth, layers in model.layers_by_depth.items():
+        for layer in layers:
+            if layer.name == 'input1':
+                input1_depth = depth
+            if layer.name == 'input2':
+                input2_depth = depth
+    assert input1_depth != -1
+    assert input1_depth == input2_depth
+
+
 if __name__ == '__main__':
     pytest.main([__file__])

tests/keras/engine/test_training.py

@@ -1,6 +1,7 @@
 import pytest
 import numpy as np
 from numpy.testing import assert_allclose
+import scipy.sparse as sparse
 
 from keras.layers import Dense, Dropout
 from keras.engine.topology import Input
@@ -198,6 +199,18 @@ def test_model_methods():
     out = model.predict([input_a_np, input_b_np], batch_size=4)
 
 
+@pytest.mark.skipif(K.backend() != 'tensorflow', reason='sparse operations supported only by TF')
+@keras_test
+def test_sparse_input_validation_split():
+    test_input = sparse.random(6, 3, density=0.25).tocsr()
+    in1 = Input(shape=(3,), sparse=True)
+    out1 = Dense(4)(in1)
+    test_output = np.random.random((6, 4))
+    model = Model(in1, out1)
+    model.compile('rmsprop', 'mse')
+    model.fit(test_input, test_output, epochs=1, batch_size=2, validation_split=0.2)
+
+
 @keras_test
 def test_trainable_argument():
     x = np.random.random((5, 3))
@@ -433,6 +446,8 @@ def test_model_with_partial_loss():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support external loss yet")
 def test_model_with_external_loss():
     # None loss, only regularization loss.
     a = Input(shape=(3,), name='input_a')

tests/keras/layers/convolutional_recurrent_test.py

@@ -3,8 +3,8 @@ import numpy as np
 from numpy.testing import assert_allclose
 
 from keras import backend as K
-from keras.models import Sequential
-from keras.layers import convolutional_recurrent
+from keras.models import Sequential, Model
+from keras.layers import convolutional_recurrent, Input
 from keras.utils.test_utils import layer_test
 from keras import regularizers
 
@@ -43,67 +43,69 @@ def test_convolutional_recurrent():
             if data_format == 'channels_first' or return_sequences:
                 continue
 
-            # Tests for statefulness
-            model = Sequential()
-            kwargs = {'data_format': data_format,
-                      'return_sequences': return_sequences,
-                      'filters': filters,
-                      'kernel_size': (num_row, num_col),
-                      'stateful': True,
-                      'batch_input_shape': inputs.shape,
-                      'padding': 'same'}
-            layer = convolutional_recurrent.ConvLSTM2D(**kwargs)
+            # cntk doesn't support statefulness on LSTM yet, will enable it on cntk later
+            if K.backend() != 'cntk':
+                # Tests for statefulness
+                model = Sequential()
+                kwargs = {'data_format': data_format,
+                          'return_sequences': return_sequences,
+                          'filters': filters,
+                          'kernel_size': (num_row, num_col),
+                          'stateful': True,
+                          'batch_input_shape': inputs.shape,
+                          'padding': 'same'}
+                layer = convolutional_recurrent.ConvLSTM2D(**kwargs)
 
-            model.add(layer)
-            model.compile(optimizer='sgd', loss='mse')
-            out1 = model.predict(np.ones_like(inputs))
+                model.add(layer)
+                model.compile(optimizer='sgd', loss='mse')
+                out1 = model.predict(np.ones_like(inputs))
 
-            # train once so that the states change
-            model.train_on_batch(np.ones_like(inputs),
-                                 np.random.random(out1.shape))
-            out2 = model.predict(np.ones_like(inputs))
+                # train once so that the states change
+                model.train_on_batch(np.ones_like(inputs),
+                                     np.random.random(out1.shape))
+                out2 = model.predict(np.ones_like(inputs))
 
-            # if the state is not reset, output should be different
-            assert(out1.max() != out2.max())
+                # if the state is not reset, output should be different
+                assert(out1.max() != out2.max())
 
-            # check that output changes after states are reset
-            # (even though the model itself didn't change)
-            layer.reset_states()
-            out3 = model.predict(np.ones_like(inputs))
-            assert(out2.max() != out3.max())
+                # check that output changes after states are reset
+                # (even though the model itself didn't change)
+                layer.reset_states()
+                out3 = model.predict(np.ones_like(inputs))
+                assert(out2.max() != out3.max())
 
-            # check that container-level reset_states() works
-            model.reset_states()
-            out4 = model.predict(np.ones_like(inputs))
-            assert_allclose(out3, out4, atol=1e-5)
+                # check that container-level reset_states() works
+                model.reset_states()
+                out4 = model.predict(np.ones_like(inputs))
+                assert_allclose(out3, out4, atol=1e-5)
 
-            # check that the call to `predict` updated the states
-            out5 = model.predict(np.ones_like(inputs))
-            assert(out4.max() != out5.max())
+                # check that the call to `predict` updated the states
+                out5 = model.predict(np.ones_like(inputs))
+                assert(out4.max() != out5.max())
 
-            # check regularizers
-            kwargs = {'data_format': data_format,
-                      'return_sequences': return_sequences,
-                      'kernel_size': (num_row, num_col),
-                      'stateful': True,
-                      'filters': filters,
-                      'batch_input_shape': inputs.shape,
-                      'kernel_regularizer': regularizers.L1L2(l1=0.01),
-                      'recurrent_regularizer': regularizers.L1L2(l1=0.01),
-                      'bias_regularizer': 'l2',
-                      'activity_regularizer': 'l2',
-                      'kernel_constraint': 'max_norm',
-                      'recurrent_constraint': 'max_norm',
-                      'bias_constraint': 'max_norm',
-                      'padding': 'same'}
+                # check regularizers
+                kwargs = {'data_format': data_format,
+                          'return_sequences': return_sequences,
+                          'kernel_size': (num_row, num_col),
+                          'stateful': True,
+                          'filters': filters,
+                          'batch_input_shape': inputs.shape,
+                          'kernel_regularizer': regularizers.L1L2(l1=0.01),
+                          'recurrent_regularizer': regularizers.L1L2(l1=0.01),
+                          'bias_regularizer': 'l2',
+                          'activity_regularizer': 'l2',
+                          'kernel_constraint': 'max_norm',
+                          'recurrent_constraint': 'max_norm',
+                          'bias_constraint': 'max_norm',
+                          'padding': 'same'}
 
-            layer = convolutional_recurrent.ConvLSTM2D(**kwargs)
-            layer.build(inputs.shape)
-            assert len(layer.losses) == 3
-            assert layer.activity_regularizer
-            output = layer(K.variable(np.ones(inputs.shape)))
-            assert len(layer.losses) == 4
-            K.eval(output)
+                layer = convolutional_recurrent.ConvLSTM2D(**kwargs)
+                layer.build(inputs.shape)
+                assert len(layer.losses) == 3
+                assert layer.activity_regularizer
+                output = layer(K.variable(np.ones(inputs.shape)))
+                assert len(layer.losses) == 4
+                K.eval(output)
 
             # check dropout
             layer_test(convolutional_recurrent.ConvLSTM2D,
@@ -116,5 +118,18 @@ def test_convolutional_recurrent():
                                'recurrent_dropout': 0.1},
                        input_shape=inputs.shape)
 
+            # check state initialization
+            layer = convolutional_recurrent.ConvLSTM2D(filters=filters,
+                                                       kernel_size=(num_row, num_col),
+                                                       data_format=data_format,
+                                                       return_sequences=return_sequences)
+            layer.build(inputs.shape)
+            x = Input(batch_shape=inputs.shape)
+            initial_state = layer.get_initial_state(x)
+            y = layer(x, initial_state=initial_state)
+            model = Model(x, y)
+            assert model.predict(inputs).shape == layer.compute_output_shape(inputs.shape)
+
+
 if __name__ == '__main__':
     pytest.main([__file__])

tests/keras/layers/convolutional_test.py

@@ -17,6 +17,8 @@ else:
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support dilated conv")
 def test_causal_dilated_conv():
     # Causal:
     layer_test(convolutional.Conv1D,
@@ -122,6 +124,8 @@ def test_averagepooling_1d():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support dilated conv")
 def test_convolution_2d():
     num_samples = 2
     filters = 2
@@ -494,20 +498,54 @@ def test_zero_padding_3d():
                      stack_size))
 
     # basic test
-    layer_test(convolutional.ZeroPadding3D,
-               kwargs={'padding': (2, 2, 2)},
-               input_shape=inputs.shape)
+    for data_format in ['channels_first', 'channels_last']:
+        layer_test(convolutional.ZeroPadding3D,
+                   kwargs={'padding': (2, 2, 2), 'data_format': data_format},
+                   input_shape=inputs.shape)
+        layer_test(convolutional.ZeroPadding3D,
+                   kwargs={'padding': ((1, 2), (3, 4), (0, 2)), 'data_format': data_format},
+                   input_shape=inputs.shape)
 
-    # correctness test
-    layer = convolutional.ZeroPadding3D(padding=(2, 2, 2))
-    layer.build(inputs.shape)
-    output = layer(K.variable(inputs))
-    np_output = K.eval(output)
-    for offset in [0, 1, -1, -2]:
-        assert_allclose(np_output[:, offset, :, :, :], 0.)
-        assert_allclose(np_output[:, :, offset, :, :], 0.)
-        assert_allclose(np_output[:, :, :, offset, :], 0.)
-    assert_allclose(np_output[:, 2:-2, 2:-2, 2:-2, :], 1.)
+        # correctness test
+        layer = convolutional.ZeroPadding3D(padding=(2, 2, 2),
+                                            data_format=data_format)
+        layer.build(inputs.shape)
+        output = layer(K.variable(inputs))
+        np_output = K.eval(output)
+        if data_format == 'channels_last':
+            for offset in [0, 1, -1, -2]:
+                assert_allclose(np_output[:, offset, :, :, :], 0.)
+                assert_allclose(np_output[:, :, offset, :, :], 0.)
+                assert_allclose(np_output[:, :, :, offset, :], 0.)
+            assert_allclose(np_output[:, 2:-2, 2:-2, 2:-2, :], 1.)
+        elif data_format == 'channels_first':
+            for offset in [0, 1, -1, -2]:
+                assert_allclose(np_output[:, :, offset, :, :], 0.)
+                assert_allclose(np_output[:, :, :, offset, :], 0.)
+                assert_allclose(np_output[:, :, :, :, offset], 0.)
+            assert_allclose(np_output[:, :, 2:-2, 2:-2, 2:-2], 1.)
+
+        layer = convolutional.ZeroPadding3D(padding=((1, 2), (3, 4), (0, 2)),
+                                            data_format=data_format)
+        layer.build(inputs.shape)
+        output = layer(K.variable(inputs))
+        np_output = K.eval(output)
+        if data_format == 'channels_last':
+            for dim1_offset in [0, -1, -2]:
+                assert_allclose(np_output[:, dim1_offset, :, :, :], 0.)
+            for dim2_offset in [0, 1, 2, -1, -2, -3, -4]:
+                assert_allclose(np_output[:, :, dim2_offset, :, :], 0.)
+            for dim3_offset in [-1, -2]:
+                assert_allclose(np_output[:, :, :, dim3_offset, :], 0.)
+            assert_allclose(np_output[:, 1:-2, 3:-4, 0:-2, :], 1.)
+        elif data_format == 'channels_first':
+            for dim1_offset in [0, -1, -2]:
+                assert_allclose(np_output[:, :, dim1_offset, :, :], 0.)
+            for dim2_offset in [0, 1, 2, -1, -2, -3, -4]:
+                assert_allclose(np_output[:, :, :, dim2_offset, :], 0.)
+            for dim3_offset in [-1, -2]:
+                assert_allclose(np_output[:, :, :, :, dim3_offset], 0.)
+            assert_allclose(np_output[:, :, 1:-2, 3:-4, 0:-2], 1.)
 
 
 @keras_test
@@ -563,6 +601,8 @@ def test_upsampling_2d():
                 assert_allclose(np_output, expected_out)
 
 
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support it yet")
 def test_upsampling_3d():
     num_samples = 2
     stack_size = 2
@@ -617,6 +657,8 @@ def test_upsampling_3d():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support slice to 0 dimension")
 def test_cropping_1d():
     num_samples = 2
     time_length = 4

tests/keras/layers/embeddings_test.py

@@ -16,6 +16,16 @@ def test_embedding():
                input_shape=(3, 2),
                input_dtype='int32',
                expected_output_dtype=K.floatx())
+    layer_test(Embedding,
+               kwargs={'output_dim': 4, 'input_dim': 10, 'mask_zero': True},
+               input_shape=(3, 2, 5),
+               input_dtype='int32',
+               expected_output_dtype=K.floatx())
+    layer_test(Embedding,
+               kwargs={'output_dim': 4, 'input_dim': 10, 'mask_zero': True, 'input_length': (None, 5)},
+               input_shape=(3, 2, 5),
+               input_dtype='int32',
+               expected_output_dtype=K.floatx())
 
 
 if __name__ == '__main__':

tests/keras/layers/noise_test.py

@@ -2,9 +2,12 @@ import pytest
 from keras.utils.test_utils import layer_test
 from keras.utils.test_utils import keras_test
 from keras.layers import noise
+from keras import backend as K
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support it yet")
 def test_GaussianNoise():
     layer_test(noise.GaussianNoise,
                kwargs={'stddev': 1.},
@@ -12,6 +15,8 @@ def test_GaussianNoise():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support it yet")
 def test_GaussianDropout():
     layer_test(noise.GaussianDropout,
                kwargs={'rate': 0.5},

tests/keras/layers/normalization_test.py

@@ -2,7 +2,7 @@ import pytest
 import numpy as np
 from numpy.testing import assert_allclose
 
-from keras.layers import Dense, Activation, Input
+from keras.layers import Input
 from keras.utils.test_utils import layer_test, keras_test
 from keras.layers import normalization
 from keras.models import Sequential, Model
@@ -52,6 +52,31 @@ def test_batchnorm_correctness():
     assert_allclose(out.std(), 1.0, atol=1e-1)
 
 
+@keras_test
+def test_batchnorm_training_argument():
+    bn1 = normalization.BatchNormalization(input_shape=(10,))
+    x1 = Input(shape=(10,))
+    y1 = bn1(x1, training=True)
+    assert bn1.updates
+
+    model1 = Model(x1, y1)
+    np.random.seed(123)
+    x = np.random.normal(loc=5.0, scale=10.0, size=(20, 10))
+    output_a = model1.predict(x)
+
+    model1.compile(loss='mse', optimizer='rmsprop')
+    model1.fit(x, x, epochs=1, verbose=0)
+    output_b = model1.predict(x)
+    assert np.abs(np.sum(output_a - output_b)) > 0.1
+    assert_allclose(output_b.mean(), 0.0, atol=1e-1)
+    assert_allclose(output_b.std(), 1.0, atol=1e-1)
+
+    bn2 = normalization.BatchNormalization(input_shape=(10,))
+    x2 = Input(shape=(10,))
+    bn2(x2, training=False)
+    assert not bn2.updates
+
+
 @keras_test
 def test_batchnorm_mode_twice():
     # This is a regression test for issue #4881 with the old

tests/keras/layers/recurrent_test.py

@@ -57,6 +57,14 @@ def test_dropout(layer_class):
                        'dropout': 0.1,
                        'recurrent_dropout': 0.1},
                input_shape=(num_samples, timesteps, embedding_dim))
+    # Test that dropout is not applied during testing
+    x = np.random.random((num_samples, timesteps, embedding_dim))
+    layer = layer_class(units, dropout=0.5, recurrent_dropout=0.5,
+                        input_shape=(timesteps, embedding_dim))
+    model = Sequential([layer])
+    y1 = model.predict(x)
+    y2 = model.predict(x)
+    assert_allclose(y1, y2)
 
 
 @rnn_test
@@ -69,6 +77,8 @@ def test_implementation_mode(layer_class):
 
 
 @rnn_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support stateful RNN yet")
 def test_statefulness(layer_class):
     model = Sequential()
     model.add(embeddings.Embedding(embedding_num, embedding_dim,
@@ -139,6 +149,8 @@ def test_regularizer(layer_class):
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support mask on RNN yet")
 def test_masking_layer():
     ''' This test based on a previously failing issue here:
     https://github.com/fchollet/keras/issues/1567
@@ -162,36 +174,50 @@ def test_masking_layer():
 
 @rnn_test
 def test_from_config(layer_class):
-    for stateful in (False, True):
+    # cntk does not support stateful yet.
+    stateful_flags = (False, True) if K.backend() != 'cntk' else (False,)
+    for stateful in stateful_flags:
         l1 = layer_class(units=1, stateful=stateful)
         l2 = layer_class.from_config(l1.get_config())
         assert l1.get_config() == l2.get_config()
 
 
 @rnn_test
-def test_specify_initial_state(layer_class):
+def test_specify_initial_state_keras_tensor(layer_class):
     num_states = 2 if layer_class is recurrent.LSTM else 1
 
     # Test with Keras tensor
     inputs = Input((timesteps, embedding_dim))
     initial_state = [Input((units,)) for _ in range(num_states)]
     layer = layer_class(units)
-    output = layer(inputs, initial_state=initial_state)
+    if len(initial_state) == 1:
+        output = layer(inputs, initial_state=initial_state[0])
+    else:
+        output = layer(inputs, initial_state=initial_state)
+    assert initial_state[0] in layer.inbound_nodes[0].input_tensors
+
     model = Model([inputs] + initial_state, output)
     model.compile(loss='categorical_crossentropy', optimizer='adam')
 
     inputs = np.random.random((num_samples, timesteps, embedding_dim))
-    initial_states = [np.random.random((num_samples, units))
-                      for _ in range(num_states)]
+    initial_state = [np.random.random((num_samples, units))
+                     for _ in range(num_states)]
     targets = np.random.random((num_samples, units))
-    model.fit([inputs] + initial_states, targets)
+    model.fit([inputs] + initial_state, targets)
+
+
+@rnn_test
+def test_specify_initial_state_non_keras_tensor(layer_class):
+    num_states = 2 if layer_class is recurrent.LSTM else 1
 
     # Test with non-Keras tensor
     inputs = Input((timesteps, embedding_dim))
-    initial_state = [K.random_normal_variable((units,), 0, 1) for _ in range(num_states)]
+    initial_state = [K.random_normal_variable((num_samples, units), 0, 1)
+                     for _ in range(num_states)]
     layer = layer_class(units)
     output = layer(inputs, initial_state=initial_state)
-    model = Model([inputs], output)
+
+    model = Model(inputs, output)
     model.compile(loss='categorical_crossentropy', optimizer='adam')
 
     inputs = np.random.random((num_samples, timesteps, embedding_dim))
@@ -200,6 +226,8 @@ def test_specify_initial_state(layer_class):
 
 
 @rnn_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support stateful RNN yet")
 def test_reset_states_with_values(layer_class):
     num_states = 2 if layer_class is recurrent.LSTM else 1
 
@@ -213,10 +241,70 @@ def test_reset_states_with_values(layer_class):
                                atol=1e-4)
     state_shapes = [K.int_shape(state) for state in layer.states]
     values = [np.ones(shape) for shape in state_shapes]
+    if len(values) == 1:
+        values = values[0]
     layer.reset_states(values)
     np.testing.assert_allclose(K.eval(layer.states[0]),
                                np.ones(K.int_shape(layer.states[0])),
                                atol=1e-4)
 
+    # Test fit with invalid data
+    with pytest.raises(ValueError):
+        layer.reset_states([1] * (len(layer.states) + 1))
+
+
+@rnn_test
+def test_specify_state_with_masking(layer_class):
+    ''' This test based on a previously failing issue here:
+    https://github.com/fchollet/keras/issues/1567
+    '''
+    num_states = 2 if layer_class is recurrent.LSTM else 1
+
+    inputs = Input((timesteps, embedding_dim))
+    _ = Masking()(inputs)
+    initial_state = [Input((units,)) for _ in range(num_states)]
+    output = layer_class(units)(inputs, initial_state=initial_state)
+
+    model = Model([inputs] + initial_state, output)
+    model.compile(loss='categorical_crossentropy', optimizer='adam')
+
+    inputs = np.random.random((num_samples, timesteps, embedding_dim))
+    initial_state = [np.random.random((num_samples, units))
+                     for _ in range(num_states)]
+    targets = np.random.random((num_samples, units))
+    model.fit([inputs] + initial_state, targets)
+
+
+@rnn_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support stateful RNN yet")
+def test_return_state(layer_class):
+    num_states = 2 if layer_class is recurrent.LSTM else 1
+
+    inputs = Input(batch_shape=(num_samples, timesteps, embedding_dim))
+    layer = layer_class(units, return_state=True, stateful=True)
+    outputs = layer(inputs)
+    output, state = outputs[0], outputs[1:]
+    assert len(state) == num_states
+    model = Model(inputs, state[0])
+
+    inputs = np.random.random((num_samples, timesteps, embedding_dim))
+    state = model.predict(inputs)
+    np.testing.assert_allclose(K.eval(layer.states[0]), state, atol=1e-4)
+
+
+@rnn_test
+def test_state_reuse(layer_class):
+    inputs = Input(batch_shape=(num_samples, timesteps, embedding_dim))
+    layer = layer_class(units, return_state=True, return_sequences=True)
+    outputs = layer(inputs)
+    output, state = outputs[0], outputs[1:]
+    output = layer_class(units)(output, initial_state=state)
+    model = Model(inputs, output)
+
+    inputs = np.random.random((num_samples, timesteps, embedding_dim))
+    outputs = model.predict(inputs)
+
+
 if __name__ == '__main__':
     pytest.main([__file__])

tests/keras/layers/wrappers_test.py

@@ -5,6 +5,7 @@ from keras.utils.test_utils import keras_test
 from keras.layers import wrappers, Input
 from keras.layers import core, convolutional, recurrent, embeddings
 from keras.models import Sequential, Model, model_from_json
+from keras import backend as K
 
 
 @keras_test
@@ -108,6 +109,8 @@ def test_regularizers():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support reverse yet")
 def test_Bidirectional():
     rnn = recurrent.SimpleRNN
     samples = 2

tests/keras/legacy/layers_test.py

@@ -184,6 +184,8 @@ def test_merge():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support stateful RNN yet")
 def test_merge_mask_2d():
     rand = lambda *shape: np.asarray(np.random.random(shape) > 0.5, dtype='int32')
 
@@ -217,6 +219,8 @@ def test_merge_mask_2d():
 
 
 @keras_test
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="cntk does not support stateful RNN yet")
 def test_merge_mask_3d():
     rand = lambda *shape: np.asarray(np.random.random(shape) > 0.5, dtype='int32')
 

tests/keras/losses_test.py

@@ -10,11 +10,14 @@ allobj = [losses.mean_squared_error,
           losses.mean_absolute_percentage_error,
           losses.mean_squared_logarithmic_error,
           losses.squared_hinge,
-          losses.hinge, losses.categorical_crossentropy,
+          losses.hinge,
+          losses.categorical_crossentropy,
           losses.binary_crossentropy,
           losses.kullback_leibler_divergence,
           losses.poisson,
-          losses.cosine_proximity]
+          losses.cosine_proximity,
+          losses.logcosh,
+          losses.categorical_hinge]
 
 
 def test_objective_shapes_3d():
@@ -44,5 +47,15 @@ def test_cce_one_hot():
     assert K.eval(losses.sparse_categorical_crossentropy(y_a, y_b)).shape == (6,)
 
 
+def test_categorical_hinge():
+    y_pred = K.variable(np.array([[0.3, 0.2, 0.1],
+                                  [0.1, 0.2, 0.7]]))
+    y_true = K.variable(np.array([[0, 1, 0],
+                                  [1, 0, 0]]))
+    expected_loss = ((0.3 - 0.2 + 1) + (0.7 - 0.1 + 1)) / 2.0
+    loss = K.eval(losses.categorical_hinge(y_true, y_pred))
+    assert np.isclose(expected_loss, np.mean(loss))
+
+
 if __name__ == '__main__':
     pytest.main([__file__])

tests/keras/metrics_test.py

@@ -17,6 +17,7 @@ all_metrics = [
     metrics.binary_crossentropy,
     metrics.poisson,
     metrics.cosine_proximity,
+    metrics.logcosh,
 ]
 
 all_sparse_metrics = [
@@ -41,6 +42,8 @@ def test_sparse_metrics():
         assert K.eval(metric(y_a, y_b)).shape == (6,)
 
 
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="keras cntk backend does not support top_k yet")
 def test_top_k_categorical_accuracy():
     y_pred = K.variable(np.array([[0.3, 0.2, 0.1], [0.1, 0.2, 0.7]]))
     y_true = K.variable(np.array([[0, 1, 0], [1, 0, 0]]))
@@ -55,5 +58,21 @@ def test_top_k_categorical_accuracy():
     assert failure_result == 0
 
 
+@pytest.mark.skipif((K.backend() == 'cntk'),
+                    reason="keras cntk backend does not support top_k yet")
+def test_sparse_top_k_categorical_accuracy():
+    y_pred = K.variable(np.array([[0.3, 0.2, 0.1], [0.1, 0.2, 0.7]]))
+    y_true = K.variable(np.array([[1], [0]]))
+    success_result = K.eval(metrics.sparse_top_k_categorical_accuracy(y_true, y_pred,
+                                                                      k=3))
+    assert success_result == 1
+    partial_result = K.eval(metrics.sparse_top_k_categorical_accuracy(y_true, y_pred,
+                                                                      k=2))
+    assert partial_result == 0.5
+    failure_result = K.eval(metrics.sparse_top_k_categorical_accuracy(y_true, y_pred,
+                                                                      k=1))
+    assert failure_result == 0
+
+
 if __name__ == '__main__':
     pytest.main([__file__])

tests/keras/preprocessing/image_test.py

@@ -73,22 +73,11 @@ class TestImage:
         with pytest.raises(ValueError):
             x = np.random.random((3, 10, 10))
             generator.fit(x)
-        with pytest.raises(ValueError):
-            x = np.random.random((32, 3, 10, 10))
-            generator.fit(x)
-        with pytest.raises(ValueError):
-            x = np.random.random((32, 10, 10, 5))
-            generator.fit(x)
+
         # Test flow with invalid data
-        with pytest.raises(ValueError):
-            x = np.random.random((32, 10, 10, 5))
-            generator.flow(np.arange(x.shape[0]))
         with pytest.raises(ValueError):
             x = np.random.random((32, 10, 10))
             generator.flow(np.arange(x.shape[0]))
-        with pytest.raises(ValueError):
-            x = np.random.random((32, 3, 10, 10))
-            generator.flow(np.arange(x.shape[0]))
 
     def test_image_data_generator_fit(self):
         generator = image.ImageDataGenerator(
@@ -160,6 +149,32 @@ class TestImage:
         assert(sorted(dir_iterator.filenames) == sorted(filenames))
         shutil.rmtree(tmp_folder)
 
+    def test_directory_iterator_class_mode_input(self):
+        tmp_folder = tempfile.mkdtemp(prefix='test_images')
+        os.mkdir(os.path.join(tmp_folder, 'class-1'))
+
+        # save the images in the paths
+        count = 0
+        for test_images in self.all_test_images:
+            for im in test_images:
+                filename = os.path.join(tmp_folder, 'class-1', 'image-{}.jpg'.format(count))
+                im.save(os.path.join(tmp_folder, filename))
+                count += 1
+
+        # create iterator
+        generator = image.ImageDataGenerator()
+        dir_iterator = generator.flow_from_directory(tmp_folder, class_mode='input')
+        batch = next(dir_iterator)
+
+        # check if input and output have the same shape
+        assert(batch[0].shape == batch[1].shape)
+        # check if the input and output images are not the same numpy array
+        input_img = batch[0][0]
+        output_img = batch[1][0]
+        output_img[0][0][0] += 1
+        assert(input_img[0][0][0] != output_img[0][0][0])
+        shutil.rmtree(tmp_folder)
+
     def test_img_utils(self):
         height, width = 10, 8
 

tests/keras/test_callbacks.py

@@ -6,9 +6,12 @@ import pytest
 from csv import Sniffer
 import shutil
 from keras import optimizers
+from keras import initializers
 from keras import callbacks
 from keras.models import Sequential
-from keras.layers.core import Dense
+from keras.layers.core import Dense, Dropout
+from keras.layers.convolutional import Conv2D
+from keras.layers.pooling import MaxPooling2D, GlobalAveragePooling2D
 from keras.utils.test_utils import get_test_data
 from keras.utils.test_utils import keras_test
 from keras import backend as K
@@ -22,6 +25,34 @@ train_samples = 20
 test_samples = 20
 
 
+@keras_test
+def test_TerminateOnNaN():
+    np.random.seed(1337)
+    (X_train, y_train), (X_test, y_test) = get_test_data(num_train=train_samples,
+                                                         num_test=test_samples,
+                                                         input_shape=(input_dim,),
+                                                         classification=True,
+                                                         num_classes=num_class)
+
+    y_test = np_utils.to_categorical(y_test)
+    y_train = np_utils.to_categorical(y_train)
+    cbks = [callbacks.TerminateOnNaN()]
+    model = Sequential()
+    initializer = initializers.Constant(value=1e5)
+    for _ in range(5):
+        model.add(Dense(num_hidden, input_dim=input_dim, activation='relu',
+                        kernel_initializer=initializer))
+    model.add(Dense(num_class, activation='linear'))
+    model.compile(loss='mean_squared_error',
+                  optimizer='rmsprop')
+
+    history = model.fit(X_train, y_train, batch_size=batch_size,
+                        validation_data=(X_test, y_test), callbacks=cbks, epochs=20)
+    loss = history.history['loss']
+    assert len(loss) == 1
+    assert loss[0] == np.inf
+
+
 @keras_test
 def test_ModelCheckpoint():
     np.random.seed(1337)
@@ -271,9 +302,9 @@ def test_CSVLogger():
 @pytest.mark.skipif((K.backend() != 'tensorflow'),
                     reason='Requires tensorflow backend')
 def test_TensorBoard():
-    np.random.seed(1337)
+    np.random.seed(np.random.randint(1, 1e7))
+    filepath = './logs_' + str(np.random.randint(1, 1e4))
 
-    filepath = './logs'
     (X_train, y_train), (X_test, y_test) = get_test_data(
         num_train=train_samples,
         num_test=test_samples,
@@ -309,13 +340,17 @@ def test_TensorBoard():
     # case 1 Sequential
     model = Sequential()
     model.add(Dense(num_hidden, input_dim=input_dim, activation='relu'))
+    model.add(Dropout(0.1))
     model.add(Dense(num_class, activation='softmax'))
     model.compile(loss='categorical_crossentropy',
                   optimizer='sgd',
                   metrics=['accuracy'])
 
     tsb = callbacks.TensorBoard(log_dir=filepath, histogram_freq=1,
-                                write_images=True)
+                                write_images=True, write_grads=True,
+                                embeddings_freq=1,
+                                embeddings_layer_names=['dense_1'],
+                                batch_size=5)
     cbks = [tsb]
 
     # fit with validation data
@@ -348,6 +383,97 @@ def test_TensorBoard():
     shutil.rmtree(filepath)
 
 
+@keras_test
+@pytest.mark.skipif((K.backend() != 'tensorflow'),
+                    reason='Requires tensorflow backend')
+def test_TensorBoard_convnet():
+    np.random.seed(np.random.randint(1, 1e7))
+    filepath = './logs_' + str(np.random.randint(1, 1e4))
+
+    input_shape = (16, 16, 3)
+    (x_train, y_train), (x_test, y_test) = get_test_data(num_train=500,
+                                                         num_test=200,
+                                                         input_shape=input_shape,
+                                                         classification=True,
+                                                         num_classes=4)
+    y_train = np_utils.to_categorical(y_train)
+    y_test = np_utils.to_categorical(y_test)
+
+    model = Sequential([
+        Conv2D(filters=8, kernel_size=3,
+               activation='relu',
+               input_shape=input_shape),
+        MaxPooling2D(pool_size=2),
+        Conv2D(filters=4, kernel_size=(3, 3),
+               activation='relu', padding='same'),
+        GlobalAveragePooling2D(),
+        Dense(y_test.shape[-1], activation='softmax')
+    ])
+    model.compile(loss='categorical_crossentropy',
+                  optimizer='rmsprop',
+                  metrics=['accuracy'])
+    tsb = callbacks.TensorBoard(log_dir=filepath, histogram_freq=1,
+                                write_images=True, write_grads=True,
+                                batch_size=16)
+    cbks = [tsb]
+    model.summary()
+    history = model.fit(x_train, y_train, epochs=2, batch_size=16,
+                        validation_data=(x_test, y_test),
+                        callbacks=cbks,
+                        verbose=0)
+    assert os.path.exists(filepath)
+    shutil.rmtree(filepath)
+
+
+@keras_test
+def test_CallbackValData():
+    np.random.seed(1337)
+    (X_train, y_train), (X_test, y_test) = get_test_data(num_train=train_samples,
+                                                         num_test=test_samples,
+                                                         input_shape=(input_dim,),
+                                                         classification=True,
+                                                         num_classes=num_class)
+    y_test = np_utils.to_categorical(y_test)
+    y_train = np_utils.to_categorical(y_train)
+    model = Sequential()
+    model.add(Dense(num_hidden, input_dim=input_dim, activation='relu'))
+    model.add(Dense(num_class, activation='softmax'))
+    model.compile(loss='categorical_crossentropy',
+                  optimizer='sgd',
+                  metrics=['accuracy'])
+
+    cbk = callbacks.LambdaCallback(on_train_end=lambda x: 1)
+    model.fit(X_train, y_train, batch_size=batch_size,
+              validation_data=(X_test, y_test), callbacks=[cbk], epochs=1)
+
+    def data_generator(train):
+        if train:
+            max_batch_index = len(X_train) // batch_size
+        else:
+            max_batch_index = len(X_test) // batch_size
+        i = 0
+        while 1:
+            if train:
+                yield (X_train[i * batch_size: (i + 1) * batch_size],
+                       y_train[i * batch_size: (i + 1) * batch_size])
+            else:
+                yield (X_test[i * batch_size: (i + 1) * batch_size],
+                       y_test[i * batch_size: (i + 1) * batch_size])
+            i += 1
+            i = i % max_batch_index
+
+    cbk2 = callbacks.LambdaCallback(on_train_end=lambda x: 1)
+    model.fit_generator(data_generator(True), len(X_train), epochs=1,
+                        validation_data=(X_test, y_test),
+                        callbacks=[cbk2])
+
+    # callback validation data should always have x, y, and sample weights
+    assert len(cbk.validation_data) == len(cbk2.validation_data) == 3
+    assert cbk.validation_data[0] is cbk2.validation_data[0]
+    assert cbk.validation_data[1] is cbk2.validation_data[1]
+    assert cbk.validation_data[2].shape == cbk2.validation_data[2].shape
+
+
 @keras_test
 def test_LambdaCallback():
     np.random.seed(1337)
@@ -386,7 +512,9 @@ def test_LambdaCallback():
                     reason="Requires tensorflow backend")
 def test_TensorBoard_with_ReduceLROnPlateau():
     import shutil
-    filepath = './logs'
+    np.random.seed(np.random.randint(1, 1e7))
+    filepath = './logs_' + str(np.random.randint(1, 1e4))
+
     (X_train, y_train), (X_test, y_test) = get_test_data(num_train=train_samples,
                                                          num_test=test_samples,
                                                          input_shape=(input_dim,),

tests/keras/utils/data_utils_test.py

@@ -9,8 +9,6 @@ from six.moves.urllib.parse import urljoin
 from keras.utils.data_utils import get_file
 from keras.utils.data_utils import validate_file
 from keras.utils.data_utils import _hash_file
-from keras import activations
-from keras import regularizers
 
 
 def test_data_utils():

tests/keras/utils/io_utils_test.py

@@ -0,0 +1,65 @@
+'''Tests for functions in io_utils.py.
+'''
+import os
+import pytest
+from keras.models import Sequential
+from keras.layers import Dense
+from keras.utils.io_utils import HDF5Matrix
+import numpy as np
+import warnings
+import h5py
+
+
+def create_dataset(h5_path='test.h5'):
+    X = np.random.randn(200, 10).astype('float32')
+    y = np.random.randint(0, 2, size=(200, 1))
+    f = h5py.File(h5_path, 'w')
+    # Creating dataset to store features
+    X_dset = f.create_dataset('my_data', (200, 10), dtype='f')
+    X_dset[:] = X
+    # Creating dataset to store labels
+    y_dset = f.create_dataset('my_labels', (200, 1), dtype='i')
+    y_dset[:] = y
+    f.close()
+
+
+def test_io_utils():
+    '''Tests the HDF5Matrix code using the sample from @jfsantos at
+    https://gist.github.com/jfsantos/e2ef822c744357a4ed16ec0c885100a3
+    '''
+    h5_path = 'test.h5'
+    create_dataset(h5_path)
+
+    # Instantiating HDF5Matrix for the training set, which is a slice of the first 150 elements
+    X_train = HDF5Matrix(h5_path, 'my_data', start=0, end=150)
+    y_train = HDF5Matrix(h5_path, 'my_labels', start=0, end=150)
+
+    # Likewise for the test set
+    X_test = HDF5Matrix(h5_path, 'my_data', start=150, end=200)
+    y_test = HDF5Matrix(h5_path, 'my_labels', start=150, end=200)
+
+    # HDF5Matrix behave more or less like Numpy matrices with regards to indexing
+    assert y_train.shape == (150, 1), 'HDF5Matrix shape should match input array'
+    # But they do not support negative indices, so don't try print(X_train[-1])
+
+    model = Sequential()
+    model.add(Dense(64, input_shape=(10,), activation='relu'))
+    model.add(Dense(1, activation='sigmoid'))
+
+    model.compile(loss='binary_crossentropy', optimizer='sgd')
+
+    # Note: you have to use shuffle='batch' or False with HDF5Matrix
+    model.fit(X_train, y_train, batch_size=32, shuffle='batch', verbose=False)
+    # test that evalutation and prediction don't crash and return reasonable results
+    out_pred = model.predict(X_test, batch_size=32, verbose=False)
+    out_eval = model.evaluate(X_test, y_test, batch_size=32, verbose=False)
+
+    assert out_pred.shape == (50, 1), 'Prediction shape does not match'
+    assert out_eval.shape == (), 'Shape of evaluation does not match'
+    assert out_eval > 0, 'Evaluation value does not meet criteria: {}'.format(out_eval)
+
+    os.remove(h5_path)
+
+
+if __name__ == '__main__':
+    pytest.main([__file__])

tests/keras/wrappers/scikit_learn_test.py

@@ -2,8 +2,6 @@ import pytest
 import numpy as np
 
 from keras.utils.test_utils import get_test_data
-from keras.utils import np_utils
-from keras import backend as K
 
 from keras.models import Sequential
 from keras.layers.core import Dense, Activation

tests/test_documentation.py

@@ -0,0 +1,151 @@
+import importlib
+import inspect
+import re
+import sys
+from itertools import compress
+
+import pytest
+
+modules = ['keras.layers', 'keras.models', 'keras', 'keras.backend.tensorflow_backend']
+accepted_name = ['from_config']
+accepted_module = ['keras.legacy.layers', 'keras.utils.generic_utils']
+
+# Functions or classes with less than 'MIN_CODE_SIZE' lines can be ignored
+MIN_CODE_SIZE = 10
+
+
+def handle_class(name, member):
+    if is_accepted(name, member):
+        return
+
+    if member.__doc__ is None and not member_too_small(member):
+        raise ValueError("{} class doesn't have any documentation".format(name),
+                         member.__module__, inspect.getmodule(member).__file__)
+    for n, met in inspect.getmembers(member):
+        if inspect.ismethod(met):
+            handle_method(n, met)
+
+
+def handle_function(name, member):
+    if is_accepted(name, member):
+        return
+    doc = member.__doc__
+    if doc is None and not member_too_small(member):
+        raise ValueError("{} function doesn't have any documentation".format(name),
+                         member.__module__, inspect.getmodule(member).__file__)
+    args = list(inspect.signature(member).parameters.keys())
+    assert_args_presence(args, doc, member, name)
+    assert_function_style(name, member, doc, args)
+    assert_doc_style(name, member, doc)
+
+
+def assert_doc_style(name, member, doc):
+    lines = doc.split("\n")
+    first_line = lines[0]
+    if len(first_line.strip()) == 0:
+        raise ValueError("{} the documentation should be on the first line.".format(name),
+                         member.__module__)
+    if first_line.strip()[-1] != '.':
+        raise ValueError("{} first line should end with a '.'".format(name),
+                         member.__module__)
+
+
+def assert_function_style(name, member, doc, args):
+    code = inspect.getsource(member)
+    has_return = re.findall(r"\s*return \S+", code, re.MULTILINE)
+    if has_return and "# Returns" not in doc:
+        innerfunction = [inspect.getsource(x) for x in member.__code__.co_consts if
+                         inspect.iscode(x)]
+        return_in_sub = [ret for code_inner in innerfunction for ret in
+                         re.findall(r"\s*return \S+", code_inner, re.MULTILINE)]
+        if len(return_in_sub) < len(has_return):
+            raise ValueError("{} needs a '# Returns' section".format(name),
+                             member.__module__)
+
+    has_raise = re.findall(r"^\s*raise \S+", code, re.MULTILINE)
+    if has_raise and "# Raises" not in doc:
+        innerfunction = [inspect.getsource(x) for x in member.__code__.co_consts if
+                         inspect.iscode(x)]
+        raise_in_sub = [ret for code_inner in innerfunction for ret in
+                        re.findall(r"\s*raise \S+", code_inner, re.MULTILINE)]
+        if len(raise_in_sub) < len(has_raise):
+            raise ValueError("{} needs a '# Raises' section".format(name),
+                             member.__module__)
+
+    if len(args) > 0 and "# Arguments" not in doc:
+        raise ValueError("{} needs a '# Arguments' section".format(name),
+                         member.__module__)
+
+    assert_blank_before(name, member, doc, ['# Arguments', '# Raises', '# Returns'])
+
+
+def assert_blank_before(name, member, doc, keywords):
+    doc_lines = [x.strip() for x in doc.split('\n')]
+    for keyword in keywords:
+        if keyword in doc_lines:
+            index = doc_lines.index(keyword)
+            if doc_lines[index - 1] != '':
+                raise ValueError(
+                    "{} '{}' should have a blank line above.".format(name, keyword),
+                    member.__module__)
+
+
+def is_accepted(name, member):
+    if 'keras' not in str(member.__module__):
+        return True
+    return name in accepted_name or member.__module__ in accepted_module
+
+
+def member_too_small(member):
+    code = inspect.getsource(member).split('\n')
+    return len(code) < MIN_CODE_SIZE
+
+
+def assert_args_presence(args, doc, member, name):
+    args_not_in_doc = [arg not in doc for arg in args]
+    if any(args_not_in_doc):
+        raise ValueError(
+            "{} {} arguments are not present in documentation ".format(name, list(
+                compress(args, args_not_in_doc))), member.__module__)
+    words = doc.replace('*', '').split()
+    # Check arguments styling
+    styles = [arg + ":" not in words for arg in args]
+    if any(styles):
+        raise ValueError(
+            "{} {} are not style properly 'argument': documentation".format(name, list(
+                compress(args, styles))), member.__module__)
+
+    # Check arguments order
+    indexes = [words.index(arg + ":") for arg in args]
+    if indexes != sorted(indexes):
+        raise ValueError(
+            "{} arguments order is different from the documentation".format(name),
+            member.__module__)
+
+
+def handle_method(name, member):
+    if name in accepted_name or member.__module__ in accepted_module:
+        return
+    handle_function(name, member)
+
+
+def handle_module(mod):
+    for name, mem in inspect.getmembers(mod):
+        if inspect.isclass(mem):
+            handle_class(name, mem)
+        elif inspect.isfunction(mem):
+            handle_function(name, mem)
+        elif 'keras' in name and inspect.ismodule(mem):
+            # Only test keras' modules
+            handle_module(mem)
+
+
+@pytest.mark.skipif(sys.version_info < (3, 3), reason="requires python3.3")
+def test_doc():
+    for module in modules:
+        mod = importlib.import_module(module)
+        handle_module(mod)
+
+
+if __name__ == '__main__':
+    pytest.main([__file__])

tests/test_model_saving.py

@@ -4,6 +4,7 @@ import tempfile
 import numpy as np
 from numpy.testing import assert_allclose
 
+from keras import backend as K
 from keras.models import Model, Sequential
 from keras.layers import Dense, Lambda, RepeatVector, TimeDistributed
 from keras.layers import Input
@@ -76,7 +77,7 @@ def test_sequential_model_saving_2():
 
 
 @keras_test
-def test_fuctional_model_saving():
+def test_functional_model_saving():
     input = Input(shape=(3,))
     x = Dense(2)(input)
     output = Dense(3)(x)
@@ -131,10 +132,11 @@ def test_saving_multiple_metrics_outputs():
 
 @keras_test
 def test_saving_without_compilation():
+    """Test saving model without compiling.
+    """
     model = Sequential()
     model.add(Dense(2, input_shape=(3,)))
     model.add(Dense(3))
-    model.compile(loss='mse', optimizer='sgd', metrics=['acc'])
 
     _, fname = tempfile.mkstemp('.h5')
     save_model(model, fname)
@@ -291,5 +293,49 @@ def test_saving_lambda_custom_objects():
     assert_allclose(out, out2, atol=1e-05)
 
 
+@keras_test
+def test_saving_lambda_numpy_array_arguments():
+    mean = np.random.random((4, 2, 3))
+    std = np.abs(np.random.random((4, 2, 3))) + 1e-5
+    input = Input(shape=(4, 2, 3))
+    output = Lambda(lambda image, mu, std: (image - mu) / std,
+                    arguments={'mu': mean, 'std': std})(input)
+    model = Model(input, output)
+    model.compile(loss='mse', optimizer='sgd', metrics=['acc'])
+
+    _, fname = tempfile.mkstemp('.h5')
+    save_model(model, fname)
+
+    model = load_model(fname)
+    os.remove(fname)
+
+    assert_allclose(mean, model.layers[1].arguments['mu'])
+    assert_allclose(std, model.layers[1].arguments['std'])
+
+
+@keras_test
+def test_saving_custom_activation_function():
+    x = Input(shape=(3,))
+    output = Dense(3, activation=K.cos)(x)
+
+    model = Model(x, output)
+    model.compile(loss=losses.MSE,
+                  optimizer=optimizers.RMSprop(lr=0.0001),
+                  metrics=[metrics.categorical_accuracy])
+    x = np.random.random((1, 3))
+    y = np.random.random((1, 3))
+    model.train_on_batch(x, y)
+
+    out = model.predict(x)
+    _, fname = tempfile.mkstemp('.h5')
+    save_model(model, fname)
+
+    model = load_model(fname, custom_objects={'cos': K.cos})
+    os.remove(fname)
+
+    out2 = model.predict(x)
+    assert_allclose(out, out2, atol=1e-05)
+
+
 if __name__ == '__main__':
     pytest.main([__file__])