Cropped Decoding¶
Now we will use cropped decoding. Cropped decoding means the ConvNet is trained on time windows/time crops within the trials. We will explain this visually by comparing trialwise to cropped decoding.
| Trialwise Decoding | Cropped Decoding |
|---|---|
 |
 |
On the left, you see trialwise decoding:
- A complete trial is pushed through the network
- The network produces a prediction
- The prediction is compared to the target (label) for that trial to compute the loss
On the right, you see cropped decoding:
- Instead of a complete trial, windows within the trial, here called crops, are pushed through the network
- For computational efficiency, multiple neighbouring crops are pushed through the network simultaneously (these neighbouring crops are called a supercrop)
- Therefore, the network produces multiple predictions (one per crop in the supercrop)
- The individual crop predictions are averaged before computing the loss function
Notes:
- The network architecture implicitly defines the crop size (it is the receptive field size, i.e., the number of timesteps the network uses to make a single prediction)
- The supercrop size is a user-defined hyperparameter, called
input_time_lengthin Braindecode. It mostly affects runtime (larger supercrop sizes should be faster). As a rule of thumb, you can set it to two times the crop size. - Crop size and supercrop size together define how many predictions the network makes per supercrop: \(\mathrm{\#supercrop}-\mathrm{\#crop}+1=\mathrm{\#predictions}\)
For cropped decoding, the above training setup is mathematically identical to sampling crops in your dataset, pushing them through the network and training directly on the individual crops. At the same time, the above training setup is much faster as it avoids redundant computations by using dilated convolutions, see our paper Deep learning with convolutional neural networks for EEG decoding and visualization. However, the two setups are only mathematically identical in case (1) your network does not use any padding and (2) your loss function leads to the same gradients when using the averaged output. The first is true for our shallow and deep ConvNet models and the second is true for the log-softmax outputs and negative log likelihood loss that is typically used for classification in PyTorch.
Most of the code for cropped decoding is identical to the Trialwise Decoding Tutorial, differences are explained in the text.
Enable logging¶
In [2]:
import logging
import importlib
importlib.reload(logging) # see https://stackoverflow.com/a/21475297/1469195
log = logging.getLogger()
log.setLevel('INFO')
import sys
logging.basicConfig(format='%(asctime)s %(levelname)s : %(message)s',
level=logging.INFO, stream=sys.stdout)
Load data¶
In [3]:
import mne
from mne.io import concatenate_raws
# 5,6,7,10,13,14 are codes for executed and imagined hands/feet
subject_id = 22
event_codes = [5,6,9,10,13,14]
#event_codes = [3,4,5,6,7,8,9,10,11,12,13,14]
# This will download the files if you don't have them yet,
# and then return the paths to the files.
physionet_paths = mne.datasets.eegbci.load_data(subject_id, event_codes)
# Load each of the files
parts = [mne.io.read_raw_edf(path, preload=True,stim_channel='auto', verbose='WARNING')
for path in physionet_paths]
# Concatenate them
raw = concatenate_raws(parts)
# Find the events in this dataset
events = mne.find_events(raw, shortest_event=0, stim_channel='STI 014')
# Use only EEG channels
eeg_channel_inds = mne.pick_types(raw.info, meg=False, eeg=True, stim=False, eog=False,
exclude='bads')
# Extract trials, only using EEG channels
epoched = mne.Epochs(raw, events, dict(hands_or_left=2, feet_or_right=3), tmin=1, tmax=4.1, proj=False, picks=eeg_channel_inds,
baseline=None, preload=True)
Convert data to Braindecode format¶
In [4]:
import numpy as np
# Convert data from volt to millivolt
# Pytorch expects float32 for input and int64 for labels.
X = (epoched.get_data() * 1e6).astype(np.float32)
y = (epoched.events[:,2] - 2).astype(np.int64) #2,3 -> 0,1
In [5]:
from braindecode.datautil.signal_target import SignalAndTarget
train_set = SignalAndTarget(X[:40], y=y[:40])
valid_set = SignalAndTarget(X[40:70], y=y[40:70])
Create the model¶
As in the trialwise decoding tutorial, we will use the Braindecode model
class directly to perform the training in a few lines of code. If you
instead want to use your own training loop, have a look at the Cropped
Manual Training Loop Tutorial.
For cropped decoding, we now transform the model into a model that
outputs a dense time series of predictions. For this, we manually set
the length of the final convolution layer to some length that makes the
receptive field of the ConvNet smaller than the number of samples in a
trial (see final_conv_length=12 in the model definition).
In [6]:
from braindecode.models.shallow_fbcsp import ShallowFBCSPNet
from torch import nn
from braindecode.torch_ext.util import set_random_seeds
# Set if you want to use GPU
# You can also use torch.cuda.is_available() to determine if cuda is available on your machine.
cuda = False
set_random_seeds(seed=20170629, cuda=cuda)
n_classes = 2
in_chans = train_set.X.shape[1]
model = ShallowFBCSPNet(in_chans=in_chans, n_classes=n_classes,
input_time_length=None,
final_conv_length=12)
if cuda:
model.cuda()
Now we supply cropped=True to our compile function
In [7]:
from braindecode.torch_ext.optimizers import AdamW
import torch.nn.functional as F
#optimizer = AdamW(model.parameters(), lr=1*0.01, weight_decay=0.5*0.001) # these are good values for the deep model
optimizer = AdamW(model.parameters(), lr=0.0625 * 0.01, weight_decay=0)
model.compile(loss=F.nll_loss, optimizer=optimizer, iterator_seed=1, cropped=True)
Run the training¶
For fitting, we must supply the super crop size. Here, we it to 450 by
setting input_time_length = 450.
In [8]:
input_time_length = 450
model.fit(train_set.X, train_set.y, epochs=30, batch_size=64, scheduler='cosine',
input_time_length=input_time_length,
validation_data=(valid_set.X, valid_set.y),)
2018-08-08 11:49:48,805 INFO : Run until first stop...
2018-08-08 11:49:52,805 INFO : Time only for training updates: 3.99s
2018-08-08 11:49:54,825 INFO : Epoch 0
2018-08-08 11:49:54,826 INFO : train_loss 4.16812
2018-08-08 11:49:54,830 INFO : valid_loss 3.39068
2018-08-08 11:49:54,830 INFO : train_misclass 0.50000
2018-08-08 11:49:54,831 INFO : valid_misclass 0.53333
2018-08-08 11:49:54,833 INFO : runtime 0.00000
2018-08-08 11:49:54,835 INFO :
2018-08-08 11:49:59,395 INFO : Time only for training updates: 4.55s
2018-08-08 11:50:03,372 INFO : Epoch 1
2018-08-08 11:50:03,379 INFO : train_loss 3.04388
2018-08-08 11:50:03,384 INFO : valid_loss 2.38477
2018-08-08 11:50:03,389 INFO : train_misclass 0.52500
2018-08-08 11:50:03,394 INFO : valid_misclass 0.53333
2018-08-08 11:50:03,398 INFO : runtime 6.59494
2018-08-08 11:50:03,403 INFO :
2018-08-08 11:50:09,317 INFO : Time only for training updates: 5.90s
2018-08-08 11:50:13,263 INFO : Epoch 2
2018-08-08 11:50:13,268 INFO : train_loss 2.06639
2018-08-08 11:50:13,273 INFO : valid_loss 1.61044
2018-08-08 11:50:13,277 INFO : train_misclass 0.47500
2018-08-08 11:50:13,281 INFO : valid_misclass 0.50000
2018-08-08 11:50:13,285 INFO : runtime 9.91786
2018-08-08 11:50:13,289 INFO :
2018-08-08 11:50:19,343 INFO : Time only for training updates: 6.03s
2018-08-08 11:50:23,245 INFO : Epoch 3
2018-08-08 11:50:23,251 INFO : train_loss 1.41936
2018-08-08 11:50:23,255 INFO : valid_loss 1.18789
2018-08-08 11:50:23,259 INFO : train_misclass 0.47500
2018-08-08 11:50:23,263 INFO : valid_misclass 0.50000
2018-08-08 11:50:23,267 INFO : runtime 10.02883
2018-08-08 11:50:23,271 INFO :
2018-08-08 11:50:29,225 INFO : Time only for training updates: 5.94s
2018-08-08 11:50:33,218 INFO : Epoch 4
2018-08-08 11:50:33,223 INFO : train_loss 1.00751
2018-08-08 11:50:33,227 INFO : valid_loss 0.95534
2018-08-08 11:50:33,232 INFO : train_misclass 0.40000
2018-08-08 11:50:33,236 INFO : valid_misclass 0.43333
2018-08-08 11:50:33,242 INFO : runtime 9.88115
2018-08-08 11:50:33,248 INFO :
2018-08-08 11:50:39,206 INFO : Time only for training updates: 5.94s
2018-08-08 11:50:43,233 INFO : Epoch 5
2018-08-08 11:50:43,237 INFO : train_loss 0.76304
2018-08-08 11:50:43,240 INFO : valid_loss 0.83303
2018-08-08 11:50:43,243 INFO : train_misclass 0.32500
2018-08-08 11:50:43,247 INFO : valid_misclass 0.43333
2018-08-08 11:50:43,250 INFO : runtime 9.97928
2018-08-08 11:50:43,255 INFO :
2018-08-08 11:50:49,142 INFO : Time only for training updates: 5.87s
2018-08-08 11:50:53,124 INFO : Epoch 6
2018-08-08 11:50:53,130 INFO : train_loss 0.61960
2018-08-08 11:50:53,134 INFO : valid_loss 0.78367
2018-08-08 11:50:53,139 INFO : train_misclass 0.30000
2018-08-08 11:50:53,143 INFO : valid_misclass 0.43333
2018-08-08 11:50:53,148 INFO : runtime 9.93809
2018-08-08 11:50:53,152 INFO :
2018-08-08 11:50:59,114 INFO : Time only for training updates: 5.95s
2018-08-08 11:51:03,191 INFO : Epoch 7
2018-08-08 11:51:03,197 INFO : train_loss 0.51544
2018-08-08 11:51:03,202 INFO : valid_loss 0.75722
2018-08-08 11:51:03,207 INFO : train_misclass 0.25000
2018-08-08 11:51:03,211 INFO : valid_misclass 0.36667
2018-08-08 11:51:03,215 INFO : runtime 9.97154
2018-08-08 11:51:03,220 INFO :
2018-08-08 11:51:09,159 INFO : Time only for training updates: 5.92s
2018-08-08 11:51:13,178 INFO : Epoch 8
2018-08-08 11:51:13,183 INFO : train_loss 0.41659
2018-08-08 11:51:13,187 INFO : valid_loss 0.70797
2018-08-08 11:51:13,191 INFO : train_misclass 0.20000
2018-08-08 11:51:13,194 INFO : valid_misclass 0.33333
2018-08-08 11:51:13,197 INFO : runtime 10.04526
2018-08-08 11:51:13,201 INFO :
2018-08-08 11:51:19,142 INFO : Time only for training updates: 5.93s
2018-08-08 11:51:23,173 INFO : Epoch 9
2018-08-08 11:51:23,195 INFO : train_loss 0.33328
2018-08-08 11:51:23,197 INFO : valid_loss 0.65720
2018-08-08 11:51:23,198 INFO : train_misclass 0.17500
2018-08-08 11:51:23,199 INFO : valid_misclass 0.30000
2018-08-08 11:51:23,200 INFO : runtime 9.98219
2018-08-08 11:51:23,201 INFO :
2018-08-08 11:51:29,120 INFO : Time only for training updates: 5.89s
2018-08-08 11:51:32,990 INFO : Epoch 10
2018-08-08 11:51:32,995 INFO : train_loss 0.27232
2018-08-08 11:51:32,999 INFO : valid_loss 0.62543
2018-08-08 11:51:33,004 INFO : train_misclass 0.15000
2018-08-08 11:51:33,008 INFO : valid_misclass 0.30000
2018-08-08 11:51:33,012 INFO : runtime 9.97608
2018-08-08 11:51:33,015 INFO :
2018-08-08 11:51:38,980 INFO : Time only for training updates: 5.95s
2018-08-08 11:51:42,894 INFO : Epoch 11
2018-08-08 11:51:42,899 INFO : train_loss 0.22401
2018-08-08 11:51:42,903 INFO : valid_loss 0.59923
2018-08-08 11:51:42,907 INFO : train_misclass 0.07500
2018-08-08 11:51:42,911 INFO : valid_misclass 0.30000
2018-08-08 11:51:42,914 INFO : runtime 9.86210
2018-08-08 11:51:42,918 INFO :
2018-08-08 11:51:48,798 INFO : Time only for training updates: 5.87s
2018-08-08 11:51:52,831 INFO : Epoch 12
2018-08-08 11:51:52,835 INFO : train_loss 0.18465
2018-08-08 11:51:52,839 INFO : valid_loss 0.56904
2018-08-08 11:51:52,843 INFO : train_misclass 0.00000
2018-08-08 11:51:52,847 INFO : valid_misclass 0.26667
2018-08-08 11:51:52,850 INFO : runtime 9.81899
2018-08-08 11:51:52,854 INFO :
2018-08-08 11:51:58,798 INFO : Time only for training updates: 5.93s
2018-08-08 11:52:02,766 INFO : Epoch 13
2018-08-08 11:52:02,771 INFO : train_loss 0.15289
2018-08-08 11:52:02,775 INFO : valid_loss 0.53387
2018-08-08 11:52:02,780 INFO : train_misclass 0.00000
2018-08-08 11:52:02,784 INFO : valid_misclass 0.23333
2018-08-08 11:52:02,788 INFO : runtime 9.99905
2018-08-08 11:52:02,793 INFO :
2018-08-08 11:52:08,751 INFO : Time only for training updates: 5.94s
2018-08-08 11:52:12,614 INFO : Epoch 14
2018-08-08 11:52:12,620 INFO : train_loss 0.12854
2018-08-08 11:52:12,627 INFO : valid_loss 0.49827
2018-08-08 11:52:12,633 INFO : train_misclass 0.00000
2018-08-08 11:52:12,637 INFO : valid_misclass 0.20000
2018-08-08 11:52:12,642 INFO : runtime 9.95297
2018-08-08 11:52:12,644 INFO :
2018-08-08 11:52:18,474 INFO : Time only for training updates: 5.82s
2018-08-08 11:52:21,438 INFO : Epoch 15
2018-08-08 11:52:21,439 INFO : train_loss 0.11070
2018-08-08 11:52:21,440 INFO : valid_loss 0.46770
2018-08-08 11:52:21,441 INFO : train_misclass 0.00000
2018-08-08 11:52:21,443 INFO : valid_misclass 0.16667
2018-08-08 11:52:21,444 INFO : runtime 9.72386
2018-08-08 11:52:21,445 INFO :
2018-08-08 11:52:24,518 INFO : Time only for training updates: 3.06s
2018-08-08 11:52:26,554 INFO : Epoch 16
2018-08-08 11:52:26,556 INFO : train_loss 0.09734
2018-08-08 11:52:26,557 INFO : valid_loss 0.44107
2018-08-08 11:52:26,558 INFO : train_misclass 0.00000
2018-08-08 11:52:26,559 INFO : valid_misclass 0.16667
2018-08-08 11:52:26,560 INFO : runtime 6.04018
2018-08-08 11:52:26,561 INFO :
2018-08-08 11:52:29,640 INFO : Time only for training updates: 3.07s
2018-08-08 11:52:31,673 INFO : Epoch 17
2018-08-08 11:52:31,674 INFO : train_loss 0.08650
2018-08-08 11:52:31,678 INFO : valid_loss 0.41685
2018-08-08 11:52:31,680 INFO : train_misclass 0.00000
2018-08-08 11:52:31,681 INFO : valid_misclass 0.16667
2018-08-08 11:52:31,682 INFO : runtime 5.12189
2018-08-08 11:52:31,683 INFO :
2018-08-08 11:52:34,754 INFO : Time only for training updates: 3.06s
2018-08-08 11:52:36,786 INFO : Epoch 18
2018-08-08 11:52:36,787 INFO : train_loss 0.07784
2018-08-08 11:52:36,791 INFO : valid_loss 0.39502
2018-08-08 11:52:36,792 INFO : train_misclass 0.00000
2018-08-08 11:52:36,793 INFO : valid_misclass 0.13333
2018-08-08 11:52:36,794 INFO : runtime 5.11526
2018-08-08 11:52:36,796 INFO :
2018-08-08 11:52:39,876 INFO : Time only for training updates: 3.07s
2018-08-08 11:52:41,910 INFO : Epoch 19
2018-08-08 11:52:41,911 INFO : train_loss 0.07096
2018-08-08 11:52:41,915 INFO : valid_loss 0.37620
2018-08-08 11:52:41,916 INFO : train_misclass 0.00000
2018-08-08 11:52:41,920 INFO : valid_misclass 0.10000
2018-08-08 11:52:41,921 INFO : runtime 5.12150
2018-08-08 11:52:41,923 INFO :
2018-08-08 11:52:45,000 INFO : Time only for training updates: 3.07s
2018-08-08 11:52:47,032 INFO : Epoch 20
2018-08-08 11:52:47,033 INFO : train_loss 0.06583
2018-08-08 11:52:47,037 INFO : valid_loss 0.36116
2018-08-08 11:52:47,038 INFO : train_misclass 0.00000
2018-08-08 11:52:47,040 INFO : valid_misclass 0.10000
2018-08-08 11:52:47,042 INFO : runtime 5.12370
2018-08-08 11:52:47,045 INFO :
2018-08-08 11:52:50,116 INFO : Time only for training updates: 3.06s
2018-08-08 11:52:52,148 INFO : Epoch 21
2018-08-08 11:52:52,149 INFO : train_loss 0.06186
2018-08-08 11:52:52,151 INFO : valid_loss 0.34902
2018-08-08 11:52:52,152 INFO : train_misclass 0.00000
2018-08-08 11:52:52,153 INFO : valid_misclass 0.10000
2018-08-08 11:52:52,154 INFO : runtime 5.11570
2018-08-08 11:52:52,156 INFO :
2018-08-08 11:52:55,236 INFO : Time only for training updates: 3.07s
2018-08-08 11:52:57,268 INFO : Epoch 22
2018-08-08 11:52:57,269 INFO : train_loss 0.05873
2018-08-08 11:52:57,273 INFO : valid_loss 0.33921
2018-08-08 11:52:57,274 INFO : train_misclass 0.00000
2018-08-08 11:52:57,278 INFO : valid_misclass 0.10000
2018-08-08 11:52:57,279 INFO : runtime 5.12058
2018-08-08 11:52:57,282 INFO :
2018-08-08 11:53:00,444 INFO : Time only for training updates: 3.15s
2018-08-08 11:53:02,601 INFO : Epoch 23
2018-08-08 11:53:02,602 INFO : train_loss 0.05626
2018-08-08 11:53:02,606 INFO : valid_loss 0.33130
2018-08-08 11:53:02,607 INFO : train_misclass 0.00000
2018-08-08 11:53:02,607 INFO : valid_misclass 0.10000
2018-08-08 11:53:02,610 INFO : runtime 5.20754
2018-08-08 11:53:02,612 INFO :
2018-08-08 11:53:05,691 INFO : Time only for training updates: 3.07s
2018-08-08 11:53:07,723 INFO : Epoch 24
2018-08-08 11:53:07,724 INFO : train_loss 0.05439
2018-08-08 11:53:07,727 INFO : valid_loss 0.32496
2018-08-08 11:53:07,728 INFO : train_misclass 0.00000
2018-08-08 11:53:07,729 INFO : valid_misclass 0.10000
2018-08-08 11:53:07,730 INFO : runtime 5.24707
2018-08-08 11:53:07,731 INFO :
2018-08-08 11:53:10,805 INFO : Time only for training updates: 3.06s
2018-08-08 11:53:12,832 INFO : Epoch 25
2018-08-08 11:53:12,833 INFO : train_loss 0.05298
2018-08-08 11:53:12,835 INFO : valid_loss 0.31987
2018-08-08 11:53:12,836 INFO : train_misclass 0.00000
2018-08-08 11:53:12,837 INFO : valid_misclass 0.10000
2018-08-08 11:53:12,838 INFO : runtime 5.11422
2018-08-08 11:53:12,839 INFO :
2018-08-08 11:53:15,915 INFO : Time only for training updates: 3.06s
2018-08-08 11:53:17,942 INFO : Epoch 26
2018-08-08 11:53:17,943 INFO : train_loss 0.05193
2018-08-08 11:53:17,947 INFO : valid_loss 0.31576
2018-08-08 11:53:17,948 INFO : train_misclass 0.00000
2018-08-08 11:53:17,949 INFO : valid_misclass 0.10000
2018-08-08 11:53:17,950 INFO : runtime 5.10983
2018-08-08 11:53:17,951 INFO :
2018-08-08 11:53:21,025 INFO : Time only for training updates: 3.06s
2018-08-08 11:53:23,051 INFO : Epoch 27
2018-08-08 11:53:23,052 INFO : train_loss 0.05121
2018-08-08 11:53:23,055 INFO : valid_loss 0.31244
2018-08-08 11:53:23,056 INFO : train_misclass 0.00000
2018-08-08 11:53:23,057 INFO : valid_misclass 0.10000
2018-08-08 11:53:23,058 INFO : runtime 5.10984
2018-08-08 11:53:23,060 INFO :
2018-08-08 11:53:26,138 INFO : Time only for training updates: 3.07s
2018-08-08 11:53:28,171 INFO : Epoch 28
2018-08-08 11:53:28,172 INFO : train_loss 0.05073
2018-08-08 11:53:28,175 INFO : valid_loss 0.30970
2018-08-08 11:53:28,176 INFO : train_misclass 0.00000
2018-08-08 11:53:28,177 INFO : valid_misclass 0.10000
2018-08-08 11:53:28,179 INFO : runtime 5.11324
2018-08-08 11:53:28,180 INFO :
2018-08-08 11:53:31,252 INFO : Time only for training updates: 3.06s
2018-08-08 11:53:33,283 INFO : Epoch 29
2018-08-08 11:53:33,284 INFO : train_loss 0.05041
2018-08-08 11:53:33,287 INFO : valid_loss 0.30740
2018-08-08 11:53:33,288 INFO : train_misclass 0.00000
2018-08-08 11:53:33,289 INFO : valid_misclass 0.10000
2018-08-08 11:53:33,290 INFO : runtime 5.11368
2018-08-08 11:53:33,292 INFO :
Out[8]:
<braindecode.experiments.experiment.Experiment at 0x7f75b82e8940>
In [9]:
model.epochs_df
Out[9]:
| train_loss | valid_loss | train_misclass | valid_misclass | runtime | |
|---|---|---|---|---|---|
| 0 | 4.168118 | 3.390682 | 0.500 | 0.533333 | 0.000000 |
| 1 | 3.043880 | 2.384769 | 0.525 | 0.533333 | 6.594939 |
| 2 | 2.066392 | 1.610444 | 0.475 | 0.500000 | 9.917859 |
| 3 | 1.419359 | 1.187893 | 0.475 | 0.500000 | 10.028832 |
| 4 | 1.007514 | 0.955341 | 0.400 | 0.433333 | 9.881147 |
| 5 | 0.763041 | 0.833030 | 0.325 | 0.433333 | 9.979284 |
| 6 | 0.619604 | 0.783671 | 0.300 | 0.433333 | 9.938087 |
| 7 | 0.515436 | 0.757223 | 0.250 | 0.366667 | 9.971540 |
| 8 | 0.416595 | 0.707971 | 0.200 | 0.333333 | 10.045257 |
| 9 | 0.333283 | 0.657200 | 0.175 | 0.300000 | 9.982185 |
| 10 | 0.272317 | 0.625433 | 0.150 | 0.300000 | 9.976075 |
| 11 | 0.224010 | 0.599227 | 0.075 | 0.300000 | 9.862099 |
| 12 | 0.184646 | 0.569042 | 0.000 | 0.266667 | 9.818995 |
| 13 | 0.152886 | 0.533874 | 0.000 | 0.233333 | 9.999050 |
| 14 | 0.128538 | 0.498271 | 0.000 | 0.200000 | 9.952966 |
| 15 | 0.110699 | 0.467696 | 0.000 | 0.166667 | 9.723856 |
| 16 | 0.097337 | 0.441067 | 0.000 | 0.166667 | 6.040177 |
| 17 | 0.086503 | 0.416848 | 0.000 | 0.166667 | 5.121891 |
| 18 | 0.077835 | 0.395015 | 0.000 | 0.133333 | 5.115263 |
| 19 | 0.070961 | 0.376196 | 0.000 | 0.100000 | 5.121497 |
| 20 | 0.065832 | 0.361156 | 0.000 | 0.100000 | 5.123702 |
| 21 | 0.061856 | 0.349016 | 0.000 | 0.100000 | 5.115699 |
| 22 | 0.058733 | 0.339215 | 0.000 | 0.100000 | 5.120580 |
| 23 | 0.056258 | 0.331300 | 0.000 | 0.100000 | 5.207545 |
| 24 | 0.054392 | 0.324961 | 0.000 | 0.100000 | 5.247069 |
| 25 | 0.052977 | 0.319866 | 0.000 | 0.100000 | 5.114219 |
| 26 | 0.051933 | 0.315760 | 0.000 | 0.100000 | 5.109835 |
| 27 | 0.051210 | 0.312438 | 0.000 | 0.100000 | 5.109844 |
| 28 | 0.050732 | 0.309702 | 0.000 | 0.100000 | 5.113238 |
| 29 | 0.050408 | 0.307397 | 0.000 | 0.100000 | 5.113682 |
Eventually, we arrive at 90% accuracy, so 27 from 30 trials are correctly predicted.
Evaluation¶
In [10]:
test_set = SignalAndTarget(X[70:], y=y[70:])
model.evaluate(test_set.X, test_set.y)
Out[10]:
{'loss': 0.4325282573699951,
'misclass': 0.09999999999999998,
'runtime': 0.0004916191101074219}
In [11]:
model.predict(test_set.X)
Out[11]:
array([1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1, 0, 1, 1, 1, 0, 1])
Dataset references¶
This dataset was created and contributed to PhysioNet by the developers of the BCI2000 instrumentation system, which they used in making these recordings. The system is described in:
Schalk, G., McFarland, D.J., Hinterberger, T., Birbaumer, N., Wolpaw, J.R. (2004) BCI2000: A General-Purpose Brain-Computer Interface (BCI) System. IEEE TBME 51(6):1034-1043.
PhysioBank is a large and growing archive of well-characterized digital recordings of physiologic signals and related data for use by the biomedical research community and further described in:
Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB, Peng C-K, Stanley HE. (2000) PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals. Circulation 101(23):e215-e220.