6398803efe
Lazy Loading ( #38 )
...
* fixing memory problems
memory problems were caused by successively appending predictions to list.
fixed by allocating memory beforehand.
* created lazy_iterators.py
added LoadCropsFromTrialsIterator
* created lazy_dataset.py
added abstract class as a parent for lazy datasets
* Update lazy_iterators.py
now saving random state before creating data loader and resetting to random state after every batch. data loader behavior was changed with pytorch upgrade 0.4.0 -> 1.0.0 and broke tests.
* Update experiment.py
* Update experiment.py
* Update lazy_iterators.py
* Update lazy_dataset.py
* Update iterators.py
* Update lazy_iterators.py
* Update iterators.py
* Update lazy_iterators.py
* Update lazy_iterators.py
added option to toggle resetting of rng state
* Update iterators.py
deactivated reset rng state after every batch by default
* Update iterators.py
accidentally changed traditional iterators. reverting changes
* Update lazy_iterators.py
deactivated resetting rng after every batch by default
* Update experiment.py
removed unused variables
2019-03-14 14:41:23 +01:00
c32ac5b0d5
fix references for AdamW
2019-01-07 15:48:44 +01:00
6c3ead2449
fixed bug in split train test
2018-10-29 20:57:35 +01:00
cab5840ecf
added padding option to avgpool2d with conv, fixed expression str in some situations
2018-10-26 17:35:43 +02:00
96ee7839b4
changed local data folder
2018-10-26 17:35:10 +02:00
651ef3d7a4
fix using cuda with model.predict
2018-09-19 12:33:34 +02:00
2760799fc0
fix reference
2018-09-19 12:32:48 +02:00
895b03703e
increment version
2018-09-17 17:38:04 +02:00
5ea159b190
adapt base model for already dense model
2018-09-17 17:23:13 +02:00
81678c4ab5
split hybrid into pytorch module and our model
2018-09-17 17:23:04 +02:00
21890f9e19
changed logic to always creting random state in function body; not in function signature
2018-09-17 17:20:21 +02:00
5f7e7fa3e4
mentioning adamw
2018-09-17 17:19:35 +02:00
d1bf8a5ca1
upgrade pytorch 0.4
2018-09-17 17:19:26 +02:00
82e6760a54
splitting into train/valid/test by folds
2018-09-17 17:18:47 +02:00
fb080fc9b9
removed seed from experiment class
2018-09-10 14:24:18 +02:00
13b791d965
updated references
2018-09-10 14:23:56 +02:00
d01feb6fa5
merge
2018-09-03 16:14:06 +02:00
b69a930799
put hyperparameters into one place in code
2018-08-27 14:22:24 +02:00
b0f186e3e0
some change to notebookdocs
2018-08-27 14:21:22 +02:00
8e82cf3db9
Merge pull request #23 from thodoris/randint_upperbound
...
Update base.py
2018-08-20 13:00:48 +02:00
ef24d89db1
Update base.py
...
changed hard-coded randit upper bound value (4294967295) with np.iinfo(np.int32).max-1.
In windows 64bit with Conda , the randint fails with the hard-coded value (4294967295).
2018-08-17 21:33:41 +03:00
bc0a83b325
logging 0 epoch now as well
2018-08-09 16:07:01 +02:00
4e27635e0b
logging 0 epoch now as well
2018-08-09 16:04:54 +02:00
c33aae643a
Braindecode 0.4, pytorch 0.4 support, simpler model class
2018-08-09 11:58:19 +02:00
c6b1fbf350
updated tests
2018-08-06 23:26:43 +02:00
b560b91dad
new simple model class
2018-08-06 23:26:31 +02:00
0da2849555
added information which criterion was triggered to Or/And
2018-08-06 23:26:12 +02:00
ea6bfe806f
pytorch 0.4 updates
2018-08-06 23:23:30 +02:00
db488aafed
added loggers, renamed compute_preds_per_trial_from_crops
2018-08-06 23:21:08 +02:00
96a62ed895
cosine annealing, cropped for basemodel, hybrid corrected
2018-07-20 15:21:17 +02:00
6f3d3c027a
merge print epoch with reset after second run
2018-07-19 17:23:07 +02:00
a3bac03277
hybrid model, eegnetv4
2018-07-19 17:18:15 +02:00
fb987d05f3
Merge pull request #22 from gemeinl/second-run-reset
...
Added resetting to best model if second run was not successful
2018-07-19 11:26:14 +02:00
cc75eed7e0
removed training whitespace
2018-07-19 09:15:40 +02:00
37de6e5bf3
Merge branch 'master' into second-run-reset
2018-07-18 15:55:54 +02:00
0bf2b597ba
added reset_after_second_run option to experiment class that defaults to false
2018-07-18 15:52:10 +02:00
cda27c6c4a
Merge pull request #20 from gemeinl/seeds
...
Adding seeding parameters to iterators and experiment class, fixing bug in resetting random state of ClassBalancedBatchSizeIterator
2018-07-18 15:33:26 +02:00
6a1c9bf588
added missing paranthesis in if statement. fixed logging of resetting to best eopch after second run
2018-07-18 14:42:49 +02:00
069e8e36d3
only resetting to epoch with best valid_misclass if valid_loss did not deceed best 1st run train_loss
2018-07-17 18:06:04 +02:00
b7f3ee055f
enabled remembering best epoch for second run after early stopping. if in the second run no valid_loss was found that deceeds the best train_loss, reset the model to the epoch with the lowest valid_misclass
2018-07-17 17:54:19 +02:00
f84b444547
fixed a bug causing the rng of ClassBalancedBatchSizeIterator to be reset to a state different from the initial random state. added seeding parameter to BalancedBatchSizeIterator and ClassBalancedBatchSizeIterator. added seeding parameter to experiment class to allow control of torch and cuda random state.
2018-07-13 11:03:03 +02:00
c4d0dea093
next version
2018-06-12 17:17:28 +02:00
fe260e2627
filtfilt option
2018-06-12 17:16:38 +02:00
0bfc5f2f3d
factored out convert angle to cart function
2018-06-12 17:16:09 +02:00
c7354c8d03
Merge branch 'master' of github.com:robintibor/braindecode
2018-06-12 17:14:08 +02:00
c3b987bbf3
rename fixed weights to pool_weights in AvgPool2dWithConv
2018-06-12 17:14:04 +02:00
ff9ae54b75
update bbci to not use dat field anymore, just nfo
2018-06-12 17:13:07 +02:00
17ee2eabd7
Merge pull request #11 from kahartma/perturbation
...
Perturbation
2018-06-08 17:11:08 +02:00
6cc7be23c9
Merge pull request #12 from kahartma/inputwindows
...
Inputwindows
2018-06-08 17:10:55 +02:00
83949a4ac7
Fixes
2018-06-08 06:46:01 +12:00
gemeinl