def get_deep_learning_model(model_args, valid_dataset):
cuda = torch.cuda.is_available()
device = model_args["device"] if cuda else 'cpu'
if cuda:
torch.backends.cudnn.benchmark = True
seed = model_args["seed"]
# = 20200220 random seed to make results reproducible
# Set random seed to be able to reproduce results
if seed:
set_random_seeds(seed=seed, cuda=cuda)
if model_args["model_type"] == "ShallowFBCSPNet":
model = ShallowFBCSPNet(
model_args["n_chans"],
model_args["n_classes"] + 1,
input_window_samples=model_args["input_window_samples"],
final_conv_length='auto',
)
elif model_args["model_type"] == "SleepStager":
model = model = SleepStager(
n_channels=model_args["n_chans"],
sfreq=model_args["sfreq"],
n_classes=model_args["n_classes"] + 1,
input_size_s=model_args["input_window_samples"] /
model_args["sfreq"],
)
else:
raise ValueError("Boom !")
if cuda:
model.cuda()
clf = EEGClassifier(
model,
criterion=model_args["criterion"],
optimizer=torch.optim.AdamW,
# using test_sample for validation
train_split=predefined_split(valid_dataset),
optimizer__lr=model_args["lr"],
optimizer__weight_decay=model_args["weight_decay"],
batch_size=model_args["batch_size"],
callbacks=[
"accuracy",
("lr_scheduler",
LRScheduler('CosineAnnealingLR',
T_max=model_args["n_epochs"] - 1)),
("early_stopping",
EarlyStopping(monitor='valid_loss',
patience=model_args["patience"]))
],
device=device,
iterator_train__num_workers=20,
iterator_train__pin_memory=True) # torch.in torch.out
return clf
def test_eeg_classifier_clonable():
preds = np.array([
[[0.2, 0.1, 0.1, 0.1], [0.8, 0.9, 0.9, 0.9]],
[[1.0, 1.0, 1.0, 1.0], [0.0, 0.0, 0.0, 0.0]],
[[1.0, 1.0, 1.0, 0.2], [0.0, 0.0, 0.0, 0.8]],
[[0.9, 0.8, 0.9, 1.0], [0.1, 0.2, 0.1, 0.0]],
])
clf = EEGClassifier(MockModule(preds),
cropped=False,
callbacks=[
"accuracy",
("lr_scheduler",
LRScheduler('CosineAnnealingLR', T_max=1))
],
criterion=CroppedLoss,
criterion__loss_function=nll_loss,
optimizer=optim.Adam,
batch_size=32)
clone(clf)
clf.initialize()
clone(clf)
def test_trialwise_predict_and_predict_proba():
preds = np.array([
[0.125, 0.875],
[1., 0.],
[0.8, 0.2],
[0.9, 0.1],
])
clf = EEGClassifier(MockModule(preds), optimizer=optim.Adam, batch_size=32)
clf.initialize()
np.testing.assert_array_equal(preds.argmax(1), clf.predict(MockDataset()))
np.testing.assert_array_equal(preds, clf.predict_proba(MockDataset()))
def test_cropped_predict_and_predict_proba_not_aggregate_predictions():
preds = np.array([
[[0.2, 0.1, 0.1, 0.1], [0.8, 0.9, 0.9, 0.9]],
[[1.0, 1.0, 1.0, 1.0], [0.0, 0.0, 0.0, 0.0]],
[[1.0, 1.0, 1.0, 0.2], [0.0, 0.0, 0.0, 0.8]],
[[0.9, 0.8, 0.9, 1.0], [0.1, 0.2, 0.1, 0.0]],
])
clf = EEGClassifier(MockModule(preds),
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=nll_loss,
optimizer=optim.Adam,
batch_size=32,
aggregate_predictions=False)
clf.initialize()
np.testing.assert_array_equal(preds.argmax(1), clf.predict(MockDataset()))
np.testing.assert_array_equal(preds, clf.predict_proba(MockDataset()))
def test_predict_trials():
preds = np.array([
[[0.2, 0.1, 0.1, 0.1], [0.8, 0.9, 0.9, 0.9]],
[[1.0, 1.0, 1.0, 1.0], [0.0, 0.0, 0.0, 0.0]],
[[1.0, 1.0, 1.0, 0.2], [0.0, 0.0, 0.0, 0.8]],
[[0.9, 0.8, 0.9, 1.0], [0.1, 0.2, 0.1, 0.0]],
])
clf = EEGClassifier(MockModule(preds),
cropped=False,
criterion=CroppedLoss,
criterion__loss_function=nll_loss,
optimizer=optim.Adam,
batch_size=32)
clf.initialize()
with pytest.warns(UserWarning,
match="This method was designed to predict "
"trials in cropped mode."):
clf.predict_trials(MockDataset(), return_targets=False)
def test_cropped_predict_and_predict_proba():
preds = np.array([
[[0.2, 0.1, 0.1, 0.1], [0.8, 0.9, 0.9, 0.9]],
[[1.0, 1.0, 1.0, 1.0], [0.0, 0.0, 0.0, 0.0]],
[[1.0, 1.0, 1.0, 0.2], [0.0, 0.0, 0.0, 0.8]],
[[0.9, 0.8, 0.9, 1.0], [0.1, 0.2, 0.1, 0.0]],
])
clf = EEGClassifier(MockModule(preds),
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=nll_loss,
optimizer=optim.Adam,
batch_size=32)
clf.initialize()
# for cropped decoding classifier returns one label for each trial (averaged over all crops)
np.testing.assert_array_equal(
preds.mean(-1).argmax(1), clf.predict(MockDataset()))
# for cropped decoding classifier returns values for each trial (average over all crops)
np.testing.assert_array_equal(preds.mean(-1),
clf.predict_proba(MockDataset()))
def exp(subject_id):
import torch
test_subj = np.r_[subject_id]
print('test subj:' + str(test_subj))
# train_subj = np.setdiff1d(np.r_[1:10], test_subj)
train_subj = np.setdiff1d(np.r_[1, 3, 7, 8], test_subj)
tr = []
val = []
for ids in train_subj:
train_size = int(0.99 * len(splitted[ids]))
test_size = len(splitted[ids]) - train_size
tr_i, val_i = torch.utils.data.random_split(splitted[ids],
[train_size, test_size])
tr.append(tr_i)
val.append(val_i)
train_set = torch.utils.data.ConcatDataset(tr)
valid_set = torch.utils.data.ConcatDataset(val)
valid_set = BaseConcatDataset([splitted[ids] for ids in test_subj])
######################################################################
# Create model
# ------------
#
######################################################################
# Now we create the deep learning model! Braindecode comes with some
# predefined convolutional neural network architectures for raw
# time-domain EEG. Here, we use the shallow ConvNet model from `Deep
# learning with convolutional neural networks for EEG decoding and
# visualization <https://arxiv.org/abs/1703.05051>`__. These models are
# pure `PyTorch <https://pytorch.org>`__ deep learning models, therefore
# to use your own model, it just has to be a normal PyTorch
# `nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__.
#
import torch
from braindecode.util import set_random_seeds
from braindecode.models import ShallowFBCSPNet, Deep4Net
cuda = torch.cuda.is_available(
) # check if GPU is available, if True chooses to use it
device = 'cuda:0' if cuda else 'cpu'
if cuda:
torch.backends.cudnn.benchmark = True
seed = 20200220 # random seed to make results reproducible
# Set random seed to be able to reproduce results
set_random_seeds(seed=seed, cuda=cuda)
n_classes = 3
# Extract number of chans and time steps from dataset
n_chans = train_set[0][0].shape[0]
input_window_samples = train_set[0][0].shape[1]
#
# model = ShallowFBCSPNet(
# n_chans,
# n_classes,
# input_window_samples=input_window_samples,
# final_conv_length='auto',
# )
from mynetworks import Deep4Net_origin, ConvClfNet, FcClfNet
model = Deep4Net(
n_chans,
n_classes,
input_window_samples=input_window_samples,
final_conv_length="auto",
)
#
# embedding_net = Deep4Net_origin(4, 22, input_window_samples)
# model = FcClfNet(embedding_net)
# #
print(model)
# Send model to GPU
if cuda:
model.cuda()
######################################################################
# Training
# --------
#
######################################################################
# Now we train the network! EEGClassifier is a Braindecode object
# responsible for managing the training of neural networks. It inherits
# from skorch.NeuralNetClassifier, so the training logic is the same as in
# `Skorch <https://skorch.readthedocs.io/en/stable/>`__.
#
######################################################################
# **Note**: In this tutorial, we use some default parameters that we
# have found to work well for motor decoding, however we strongly
# encourage you to perform your own hyperparameter optimization using
# cross validation on your training data.
#
from skorch.callbacks import LRScheduler
from skorch.helper import predefined_split
from braindecode import EEGClassifier
# # These values we found good for shallow network:
lr = 0.0625 * 0.01
weight_decay = 0
# For deep4 they should be:
# lr = 1 * 0.01
# weight_decay = 0.5 * 0.001
batch_size = 8
n_epochs = 100
clf = EEGClassifier(
model,
criterion=torch.nn.NLLLoss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(
valid_set), # using valid_set for validation
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
batch_size=batch_size,
callbacks=[
"accuracy",
("lr_scheduler",
LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Plot Results
# ------------
#
######################################################################
# Now we use the history stored by Skorch throughout training to plot
# accuracy and loss curves.
#
import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
import pandas as pd
# Extract loss and accuracy values for plotting from history object
results_columns = [
'train_loss', 'valid_loss', 'train_accuracy', 'valid_accuracy'
]
df = pd.DataFrame(clf.history[:, results_columns],
columns=results_columns,
index=clf.history[:, 'epoch'])
# get percent of misclass for better visual comparison to loss
df = df.assign(train_misclass=100 - 100 * df.train_accuracy,
valid_misclass=100 - 100 * df.valid_accuracy)
plt.style.use('seaborn')
fig, ax1 = plt.subplots(figsize=(8, 3))
df.loc[:, ['train_loss', 'valid_loss']].plot(ax=ax1,
style=['-', ':'],
marker='o',
color='tab:blue',
legend=False,
fontsize=14)
ax1.tick_params(axis='y', labelcolor='tab:blue', labelsize=14)
ax1.set_ylabel("Loss", color='tab:blue', fontsize=14)
ax2 = ax1.twinx() # instantiate a second axes that shares the same x-axis
df.loc[:, ['train_misclass', 'valid_misclass']].plot(ax=ax2,
style=['-', ':'],
marker='o',
color='tab:red',
legend=False)
ax2.tick_params(axis='y', labelcolor='tab:red', labelsize=14)
ax2.set_ylabel("Misclassification Rate [%]", color='tab:red', fontsize=14)
ax2.set_ylim(ax2.get_ylim()[0], 85) # make some room for legend
ax1.set_xlabel("Epoch", fontsize=14)
# where some data has already been plotted to ax
handles = []
handles.append(
Line2D([0], [0],
color='black',
linewidth=1,
linestyle='-',
label='Train'))
handles.append(
Line2D([0], [0],
color='black',
linewidth=1,
linestyle=':',
label='Valid'))
plt.legend(handles, [h.get_label() for h in handles], fontsize=14)
plt.tight_layout()
# plt.show()
return df
train_bal_acc = EpochScoring(scoring='balanced_accuracy',
on_train=True,
name='train_bal_acc',
lower_is_better=False)
valid_bal_acc = EpochScoring(scoring='balanced_accuracy',
on_train=False,
name='valid_bal_acc',
lower_is_better=False)
callbacks = [('train_bal_acc', train_bal_acc),
('valid_bal_acc', valid_bal_acc)]
clf = EEGClassifier(
model,
criterion=torch.nn.CrossEntropyLoss,
optimizer=torch.optim.Adam,
train_split=predefined_split(valid_set), # using valid_set for validation
optimizer__lr=lr,
batch_size=batch_size,
callbacks=callbacks,
device=device)
# Model training for a specified number of epochs. `y` is None as it is already
# supplied in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Plot results
# ------------
#
######################################################################
# We use the history stored by Skorch during training to plot the performance of
if re.compile('/Users/long/').match(location):
my_callbacks = [
"accuracy",
("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
('on_epoch_begin_callback', on_epoch_begin_callback),
('on_batch_end_callback', on_batch_end_callback),
]
elif re.compile('/content/drive').match(location):
my_callbacks = [
"accuracy",
("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
]
clf = EEGClassifier(
net,
#criterion=torch.nn.NLLLoss, #torch.nn.NLLLoss/CrossEntropyLoss
criterion=torch.nn.CrossEntropyLoss,
optimizer=torch.optim.Adam, #optimizer=torch.optim.AdamW,
train_split=predefined_split(
valid_set
), # using valid_set for validation; None means no validate:both train and test on training dataset.
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
batch_size=batch_size,
callbacks=my_callbacks,
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
batch_size = 64
n_epochs = 100
from skorch.callbacks import Checkpoint, EarlyStopping, EpochScoring
early_stopping = EarlyStopping(patience=30)
cp = Checkpoint(dirname='', f_criterion=None, f_optimizer=None, f_history=None)
clf = EEGClassifier(
model,
# criterion=torch.nn.NLLLoss,
criterion=torch.nn.CrossEntropyLoss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(valid_set), # using valid_set for validation
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
batch_size=batch_size,
callbacks=[
"accuracy", ("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),('cp', cp),('patience', early_stopping)
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
clf.load_params(checkpoint=cp) # Load the model with the lowest valid_loss
import os
os.remove('./params.pt') # Delete parameters file
######################################################################
def test_variable_length_trials_cropped_decoding():
cuda = False
set_random_seeds(seed=20210726, cuda=cuda)
# create fake tuh abnormal dataset
tuh = _TUHAbnormalMock(path='')
# fake variable length trials by cropping first recording
splits = tuh.split([[i] for i in range(len(tuh.datasets))])
preprocess(
concat_ds=splits['0'],
preprocessors=[
Preprocessor('crop', tmax=300),
],
)
variable_tuh = BaseConcatDataset(
[splits[str(i)] for i in range(len(tuh.datasets))])
# make sure we actually have different length trials
assert any(np.diff([ds.raw.n_times for ds in variable_tuh.datasets]) != 0)
# create windows
variable_tuh_windows = create_fixed_length_windows(
concat_ds=variable_tuh,
window_size_samples=1000,
window_stride_samples=1000,
drop_last_window=False,
mapping={
True: 1,
False: 0
},
)
# create train and valid set
splits = variable_tuh_windows.split(
[[i] for i in range(len(variable_tuh_windows.datasets))])
variable_tuh_windows_train = BaseConcatDataset(
[splits[str(i)] for i in range(len(tuh.datasets) - 1)])
variable_tuh_windows_valid = BaseConcatDataset(
[splits[str(len(tuh.datasets) - 1)]])
for x, y, ind in variable_tuh_windows_train:
break
train_split = predefined_split(variable_tuh_windows_valid)
# initialize a model
model = ShallowFBCSPNet(
in_chans=x.shape[0],
n_classes=len(tuh.description.pathological.unique()),
)
to_dense_prediction_model(model)
if cuda:
model.cuda()
# create and train a classifier
clf = EEGClassifier(
model,
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=torch.nn.functional.nll_loss,
optimizer=torch.optim.Adam,
batch_size=32,
callbacks=['accuracy'],
train_split=train_split,
)
clf.fit(variable_tuh_windows_train, y=None, epochs=3)
# make sure it does what we expect
np.testing.assert_allclose(
clf.history[:, 'train_loss'],
np.array([
0.689495325088501,
0.1353449523448944,
0.006638816092163324,
]),
rtol=1e-1,
atol=1e-1,
)
np.testing.assert_allclose(
clf.history[:, 'valid_loss'],
np.array([
2.925871,
3.611423,
4.23494,
]),
rtol=1e-1,
atol=1e-1,
)
args.input_window_samples)[2] valid_set, train_set = windowing_data(dataset, args) del dataset # ============================================================================= # Training the model # ============================================================================= clf = EEGClassifier( model, cropped=True, criterion=CroppedLoss, criterion__loss_function=torch.nn.functional.nll_loss, optimizer=torch.optim.AdamW, train_split=predefined_split(valid_set), optimizer__lr=args.lr, optimizer__weight_decay=args.wd, iterator_train__shuffle=True, batch_size=args.batch_size, callbacks=['accuracy'], device=args.device, ) # Model training for a specified number of epochs. "y" is None as it is already # supplied in the dataset. clf.fit(train_set, y=None, epochs=args.n_epochs) # ============================================================================= # Storing the results into a dictionary and saving # =============================================================================
name='train_acc',
lower_is_better=False)
valid_acc = EpochScoring(scoring='accuracy',
on_train=False,
name='valid_acc',
lower_is_better=False)
callbacks = [('cp', cp), ('patience', early_stopping),
('train_acc', train_acc), ('valid_acc', valid_acc)]
clf = EEGClassifier(model,
criterion=torch.nn.BCEWithLogitsLoss,
optimizer=torch.optim.Adam,
max_epochs=n_epochs,
iterator_train__shuffle=False,
iterator_train__sampler=train_sampler,
iterator_valid__sampler=valid_sampler,
iterator_train__num_workers=num_workers,
iterator_valid__num_workers=num_workers,
train_split=predefined_split(splitted['valid']),
optimizer__lr=lr,
batch_size=batch_size,
callbacks=callbacks,
device=device)
# Model training for a specified number of epochs. `y` is None as it is already
# supplied in the dataset.
clf.fit(splitted['train'], y=None)
clf.load_params(checkpoint=cp) # Load the model with the lowest valid_loss
os.remove('./params.pt') # Delete parameters file
######################################################################
# Visualizing the results
preload=True,
)
splitted = windows_dataset.split('session')
train_set = splitted['session_T']
valid_set = splitted['session_E']
clf = EEGClassifier(
model,
cropped=False,
criterion=torch.nn.NLLLoss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(valid_set),
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
iterator_train__shuffle=True,
batch_size=batch_size,
callbacks=[
"accuracy",
# seems n_epochs -1 leads to desired behavior of lr=0 after end of training?
("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
clf.fit(windows_dataset, y=None, epochs=n_epochs)
###############################################################################
# Plot Results
ignore_keys = [
def exp(subject_id):
test_subj = np.r_[subject_id]
print('test subj:' + str(test_subj))
train_subj = np.setdiff1d(np.r_[1:11], test_subj)
train_set = BaseConcatDataset([splitted[ids] for ids in train_subj])
valid_set = BaseConcatDataset([splitted[ids] for ids in test_subj])
# #
model = ShallowFBCSPNet(
n_chans,
n_classes,
input_window_samples=input_window_samples,
final_conv_length=30,
)
# #
# embedding_net = Deep4Net_origin(4, 22, input_window_samples)
# model = FcClfNet(embedding_net)
print(model)
# Send model to GPU
if cuda:
model.cuda()
from braindecode.models.util import to_dense_prediction_model, get_output_shape
to_dense_prediction_model(model)
######################################################################
# Training
# --------
#
######################################################################
# In difference to trialwise decoding, we now should supply
# ``cropped=True`` to the EEGClassifier, and ``CroppedLoss`` as the
# criterion, as well as ``criterion__loss_function`` as the loss function
# applied to the meaned predictions.
#
######################################################################
# .. note::
# In this tutorial, we use some default parameters that we
# have found to work well for motor decoding, however we strongly
# encourage you to perform your own hyperparameter optimization using
# cross validation on your training data.
#
from skorch.callbacks import LRScheduler
from skorch.helper import predefined_split
from braindecode import EEGClassifier
from braindecode.training.losses import CroppedLoss
from braindecode.training.scoring import trial_preds_from_window_preds
# # These values we found good for shallow network:
lr = 0.0625 * 0.01
weight_decay = 0
# # For deep4 they should be:
# lr = 1 * 0.01
# weight_decay = 0.5 * 0.001
batch_size = 400
n_epochs = 100
clf = EEGClassifier(
model,
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=torch.nn.functional.nll_loss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(valid_set),
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
iterator_train__shuffle=True,
batch_size=batch_size,
callbacks=[
"accuracy",
("lr_scheduler",
LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Plot Results
# ------------
#
######################################################################
# This is again the same code as in trialwise decoding.
#
# .. note::
# Note that we drop further in the classification error and
# loss as in the trialwise decoding tutorial.
#
import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
import pandas as pd
# Extract loss and accuracy values for plotting from history object
results_columns = [
'train_loss', 'valid_loss', 'train_accuracy', 'valid_accuracy'
]
df = pd.DataFrame(clf.history[:, results_columns],
columns=results_columns,
index=clf.history[:, 'epoch'])
# get percent of misclass for better visual comparison to loss
df = df.assign(train_misclass=100 - 100 * df.train_accuracy,
valid_misclass=100 - 100 * df.valid_accuracy)
plt.style.use('seaborn')
fig, ax1 = plt.subplots(figsize=(8, 3))
df.loc[:, ['train_loss', 'valid_loss']].plot(ax=ax1,
style=['-', ':'],
marker='o',
color='tab:blue',
legend=False,
fontsize=14)
ax1.tick_params(axis='y', labelcolor='tab:blue', labelsize=14)
ax1.set_ylabel("Loss", color='tab:blue', fontsize=14)
ax2 = ax1.twinx() # instantiate a second axes that shares the same x-axis
df.loc[:, ['train_misclass', 'valid_misclass']].plot(ax=ax2,
style=['-', ':'],
marker='o',
color='tab:red',
legend=False)
ax2.tick_params(axis='y', labelcolor='tab:red', labelsize=14)
ax2.set_ylabel("Misclassification Rate [%]", color='tab:red', fontsize=14)
ax2.set_ylim(ax2.get_ylim()[0], 85) # make some room for legend
ax1.set_xlabel("Epoch", fontsize=14)
# where some data has already been plotted to ax
handles = []
handles.append(
Line2D([0], [0],
color='black',
linewidth=1,
linestyle='-',
label='Train'))
handles.append(
Line2D([0], [0],
color='black',
linewidth=1,
linestyle=':',
label='Valid'))
plt.legend(handles, [h.get_label() for h in handles], fontsize=14)
plt.tight_layout()
return df
def exp(subject_id):
dataset = MOABBDataset(dataset_name="BNCI2014001", subject_ids=subject_id)
from braindecode.datautil.preprocess import exponential_moving_standardize
from braindecode.datautil.preprocess import MNEPreproc, NumpyPreproc, preprocess
low_cut_hz = 0. # low cut frequency for filtering
high_cut_hz = 38. # high cut frequency for filtering
# Parameters for exponential moving standardization
factor_new = 1e-3
init_block_size = 1000
preprocessors = [
# keep only EEG sensors
MNEPreproc(fn='pick_types', eeg=True, meg=False, stim=False),
# convert from volt to microvolt, directly modifying the numpy array
NumpyPreproc(fn=lambda x: x * 1e6),
# bandpass filter
MNEPreproc(fn='filter', l_freq=low_cut_hz, h_freq=high_cut_hz),
# exponential moving standardization
# NumpyPreproc(fn=exponential_moving_standardize, factor_new=factor_new,
# init_block_size=init_block_size)
]
# Transform the data
preprocess(dataset, preprocessors)
######################################################################
# Create model and compute windowing parameters
# ---------------------------------------------
#
######################################################################
# In contrast to trialwise decoding, we first have to create the model
# before we can cut the dataset into windows. This is because we need to
# know the receptive field of the network to know how large the window
# stride should be.
#
######################################################################
# We first choose the compute/input window size that will be fed to the
# network during training This has to be larger than the networks
# receptive field size and can otherwise be chosen for computational
# efficiency (see explanations in the beginning of this tutorial). Here we
# choose 1000 samples, which are 4 seconds for the 250 Hz sampling rate.
#
input_window_samples = 1000
######################################################################
# Now we create the model. To enable it to be used in cropped decoding
# efficiently, we manually set the length of the final convolution layer
# to some length that makes the receptive field of the ConvNet smaller
# than ``input_window_samples`` (see ``final_conv_length=30`` in the model
# definition).
#
import torch
from braindecode.util import set_random_seeds
from braindecode.models import ShallowFBCSPNet, Deep4Net
cuda = torch.cuda.is_available(
) # check if GPU is available, if True chooses to use it
device = 'cuda:1' if cuda else 'cpu'
if cuda:
torch.backends.cudnn.benchmark = True
seed = 20190706 # random seed to make results reproducible
# Set random seed to be able to reproduce results
set_random_seeds(seed=seed, cuda=cuda)
n_classes = 4
# Extract number of chans from dataset
n_chans = dataset[0][0].shape[0]
# model = Deep4Net(
# n_chans,
# n_classes,
# input_window_samples=input_window_samples,
# final_conv_length="auto",
# )
#
#
#
# embedding_net = Deep4Net_origin(4, 22, input_window_samples)
# model = FcClfNet(embedding_net)
model = ShallowFBCSPNet(
n_chans,
n_classes,
input_window_samples=input_window_samples,
final_conv_length=30,
)
print(model)
# Send model to GPU
if cuda:
model.cuda(device)
######################################################################
# And now we transform model with strides to a model that outputs dense
# prediction, so we can use it to obtain predictions for all
# crops.
#
from braindecode.models.util import to_dense_prediction_model, get_output_shape
to_dense_prediction_model(model)
n_preds_per_input = get_output_shape(model, n_chans,
input_window_samples)[2]
print("n_preds_per_input : ", n_preds_per_input)
print(model)
######################################################################
# Cut the data into windows
# -------------------------
#
######################################################################
# In contrast to trialwise decoding, we have to supply an explicit window size and window stride to the
# ``create_windows_from_events`` function.
#
import numpy as np
from braindecode.datautil.windowers import create_windows_from_events
trial_start_offset_seconds = -0.5
# Extract sampling frequency, check that they are same in all datasets
sfreq = dataset.datasets[0].raw.info['sfreq']
assert all([ds.raw.info['sfreq'] == sfreq for ds in dataset.datasets])
# Calculate the trial start offset in samples.
trial_start_offset_samples = int(trial_start_offset_seconds * sfreq)
# Create windows using braindecode function for this. It needs parameters to define how
# trials should be used.
windows_dataset = create_windows_from_events(
dataset,
trial_start_offset_samples=trial_start_offset_samples,
trial_stop_offset_samples=0,
window_size_samples=input_window_samples,
window_stride_samples=n_preds_per_input,
drop_last_window=False,
preload=True,
)
######################################################################
# Split the dataset
# -----------------
#
# This code is the same as in trialwise decoding.
#
from braindecode.datasets.base import BaseConcatDataset
splitted = windows_dataset.split('session')
train_set = splitted['session_T']
valid_set = splitted['session_E']
######################################################################
# In difference to trialwise decoding, we now should supply
# ``cropped=True`` to the EEGClassifier, and ``CroppedLoss`` as the
# criterion, as well as ``criterion__loss_function`` as the loss function
# applied to the meaned predictions.
#
######################################################################
# .. note::
# In this tutorial, we use some default parameters that we
# have found to work well for motor decoding, however we strongly
# encourage you to perform your own hyperparameter optimization using
# cross validation on your training data.
#
from skorch.callbacks import LRScheduler
from skorch.helper import predefined_split
from braindecode import EEGClassifier
from braindecode.training.losses import CroppedLoss
from braindecode.training.scoring import trial_preds_from_window_preds
# # These values we found good for shallow network:
lr = 0.0625 * 0.01
weight_decay = 0
# # For deep4 they should be:
# lr = 1 * 0.01
# weight_decay = 0.5 * 0.001
#
batch_size = 8
n_epochs = 100
clf = EEGClassifier(
model,
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=torch.nn.functional.nll_loss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(valid_set),
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
iterator_train__shuffle=True,
batch_size=batch_size,
callbacks=[
"accuracy",
("lr_scheduler",
LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Plot Results
# ------------
#
######################################################################
# This is again the same code as in trialwise decoding.
#
# .. note::
# Note that we drop further in the classification error and
# loss as in the trialwise decoding tutorial.
#
import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
import pandas as pd
# Extract loss and accuracy values for plotting from history object
results_columns = [
'train_loss', 'valid_loss', 'train_accuracy', 'valid_accuracy'
]
df = pd.DataFrame(clf.history[:, results_columns],
columns=results_columns,
index=clf.history[:, 'epoch'])
# get percent of misclass for better visual comparison to loss
df = df.assign(train_misclass=100 - 100 * df.train_accuracy,
valid_misclass=100 - 100 * df.valid_accuracy)
plt.style.use('seaborn')
fig, ax1 = plt.subplots(figsize=(8, 3))
df.loc[:, ['train_loss', 'valid_loss']].plot(ax=ax1,
style=['-', ':'],
marker='o',
color='tab:blue',
legend=False,
fontsize=14)
ax1.tick_params(axis='y', labelcolor='tab:blue', labelsize=14)
ax1.set_ylabel("Loss", color='tab:blue', fontsize=14)
ax2 = ax1.twinx() # instantiate a second axes that shares the same x-axis
df.loc[:, ['train_misclass', 'valid_misclass']].plot(ax=ax2,
style=['-', ':'],
marker='o',
color='tab:red',
legend=False)
ax2.tick_params(axis='y', labelcolor='tab:red', labelsize=14)
ax2.set_ylabel("Misclassification Rate [%]", color='tab:red', fontsize=14)
ax2.set_ylim(ax2.get_ylim()[0], 85) # make some room for legend
ax1.set_xlabel("Epoch", fontsize=14)
# where some data has already been plotted to ax
handles = []
handles.append(
Line2D([0], [0],
color='black',
linewidth=1,
linestyle='-',
label='Train'))
handles.append(
Line2D([0], [0],
color='black',
linewidth=1,
linestyle=':',
label='Valid'))
plt.legend(handles, [h.get_label() for h in handles], fontsize=14)
plt.tight_layout()
return df
# For deep4 they should be:
# lr = 1 * 0.01
# weight_decay = 0.5 * 0.001
batch_size = 64
n_epochs = 4
clf = EEGClassifier(
model,
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=torch.nn.functional.nll_loss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(valid_set),
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
iterator_train__shuffle=True,
batch_size=batch_size,
callbacks=[
"accuracy", ("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Plot Results
# ------------
lr = 0.0625 * 0.01
weight_decay = 0
batch_size = 64
n_epochs = 4
clf = EEGClassifier(
model,
iterator_train=
AugmentedDataLoader, # This tells EEGClassifier to use a custom DataLoader
iterator_train__transforms=transforms, # This sets the augmentations to use
criterion=torch.nn.NLLLoss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(valid_set), # using valid_set for validation
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
batch_size=batch_size,
callbacks=[
"accuracy",
("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already
# supplied in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Manually composing Transforms
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
def test_cropped_decoding():
# 5,6,7,10,13,14 are codes for executed and imagined hands/feet
subject_id = 1
event_codes = [5, 6, 9, 10, 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, update_path=False
)
# 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.events_from_annotations(raw)
# 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=2, feet=3),
tmin=1,
tmax=4.1,
proj=False,
picks=eeg_channel_inds,
baseline=None,
preload=True,
)
# 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
# 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)
# This will determine how many crops are processed in parallel
input_time_length = 450
n_classes = 2
in_chans = X.shape[1]
# final_conv_length determines the size of the receptive field of the ConvNet
model = ShallowFBCSPNet(
in_chans=in_chans,
n_classes=n_classes,
input_time_length=input_time_length,
final_conv_length=12,
)
to_dense_prediction_model(model)
if cuda:
model.cuda()
# Perform forward pass to determine how many outputs per input
n_preds_per_input = get_output_shape(model, in_chans, input_time_length)[2]
train_set = CroppedXyDataset(X[:60], y[:60],
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
valid_set = CroppedXyDataset(X[60:], y=y[60:],
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
train_split = predefined_split(valid_set)
clf = EEGClassifier(
model,
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=torch.nn.functional.nll_loss,
optimizer=optim.Adam,
train_split=train_split,
batch_size=32,
callbacks=['accuracy'],
)
clf.fit(train_set, y=None, epochs=4)
np.testing.assert_allclose(
clf.history[:, 'train_loss'],
np.array(
[
1.455306,
1.455934,
1.210563,
1.065806
]
),
rtol=1e-4,
atol=1e-5,
)
np.testing.assert_allclose(
clf.history[:, 'valid_loss'],
np.array(
[
2.547288,
1.51785,
1.394036,
1.064355
]
),
rtol=1e-4,
atol=1e-4,
)
np.testing.assert_allclose(
clf.history[:, 'train_accuracy'],
np.array(
[
0.5,
0.5,
0.5,
0.533333
]
),
rtol=1e-4,
atol=1e-5,
)
np.testing.assert_allclose(
clf.history[:, 'valid_accuracy'],
np.array(
[
0.533333,
0.466667,
0.533333,
0.5
]
),
rtol=1e-4,
atol=1e-5,
)
weight_decay = 0
# For deep4 they should be:
# lr = 1 * 0.01
# weight_decay = 0.5 * 0.001
batch_size = 64
n_epochs = 4
clf = EEGClassifier(
model,
criterion=torch.nn.NLLLoss,
optimizer=torch.optim.AdamW,
train_split=predefined_split(valid_set), # using valid_set for validation
optimizer__lr=lr,
optimizer__weight_decay=weight_decay,
batch_size=batch_size,
callbacks=[
"accuracy",
("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Plot Results
# ------------
#
name='train_bal_acc',
lower_is_better=False)
valid_bal_acc = EpochScoring(scoring='balanced_accuracy',
on_train=False,
name='valid_bal_acc',
lower_is_better=False)
callbacks = [('train_bal_acc', train_bal_acc),
('valid_bal_acc', valid_bal_acc)]
clf = EEGClassifier(
model,
criterion=torch.nn.CrossEntropyLoss,
criterion__weight=torch.Tensor(class_weights).to(device),
optimizer=torch.optim.Adam,
iterator_train__shuffle=False,
iterator_train__sampler=train_sampler,
iterator_valid__sampler=valid_sampler,
train_split=predefined_split(valid_set), # using valid_set for validation
optimizer__lr=lr,
batch_size=batch_size,
callbacks=callbacks,
device=device)
# Model training for a specified number of epochs. `y` is None as it is already
# supplied in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
######################################################################
# Plot results
# ------------
#
# We use the history stored by Skorch during training to plot the performance of
def train(subject_id):
subject_range = [subject_id]
##### subject_range = [x for x in range(1, 10)]
dataset = MOABBDataset(dataset_name="BNCI2014001",
subject_ids=subject_range)
######################################################################
# Preprocessing
low_cut_hz = 4. # low cut frequency for filtering
high_cut_hz = 38. # high cut frequency for filtering
# Parameters for exponential moving standardization
factor_new = 1e-3
init_block_size = 1000
preprocessors = [
Preprocessor('pick_types', eeg=True, eog=False, meg=False,
stim=False), # Keep EEG sensors
Preprocessor(lambda x: x * 1e6), # Convert from V to uV
Preprocessor('filter', l_freq=low_cut_hz,
h_freq=high_cut_hz), # Bandpass filter
Preprocessor('set_eeg_reference',
ref_channels='average',
ch_type='eeg'),
Preprocessor('resample', sfreq=125),
## Preprocessor(exponential_moving_standardize, # Exponential moving standardization
## factor_new=factor_new, init_block_size=init_block_size)
## Preprocessor('pick_channels', ch_names=short_ch_names, ordered=True),
]
# Transform the data
preprocess(dataset, preprocessors)
######################################################################
# Cut Compute Windows
# ~~~~~~~~~~~~~~~~~~~
trial_start_offset_seconds = -0.0
# Extract sampling frequency, check that they are same in all datasets
sfreq = dataset.datasets[0].raw.info['sfreq']
assert all([ds.raw.info['sfreq'] == sfreq for ds in dataset.datasets])
# Calculate the trial start offset in samples.
trial_start_offset_samples = int(trial_start_offset_seconds * sfreq)
# Create windows using braindecode function for this. It needs parameters to define how
# trials should be used.
windows_dataset = create_windows_from_events(
dataset,
# picks=["Fz", "FC3", "FC1", "FCz", "FC2", "FC4", "C5", "C3", "C1", "Cz", "C2", "C4", "C6", "CP3", "CP1", "CPz", "CP2", "CP4", "P1", "Pz", "P2", "POz"],
trial_start_offset_samples=trial_start_offset_samples,
trial_stop_offset_samples=0,
preload=True,
)
######################################################################
# Split dataset into train and valid
splitted = windows_dataset.split('session')
train_set = splitted['session_T']
valid_set = splitted['session_E']
######################################################################
# Create model
cuda = torch.cuda.is_available(
) # check if GPU is available, if True chooses to use it
device = 'cuda' if cuda else 'cpu'
if cuda:
torch.backends.cudnn.benchmark = True
seed = 20200220 # random seed to make results reproducible
# Set random seed to be able to reproduce results
set_random_seeds(seed=seed, cuda=cuda)
n_classes = 4
# Extract number of chans and time steps from dataset
n_chans = train_set[0][0].shape[0]
input_window_samples = train_set[0][0].shape[1]
model = ShallowFBCSPNet(n_chans,
n_classes,
input_window_samples=input_window_samples,
final_conv_length='auto')
"""
model = EEGNetv1(
n_chans,
n_classes,
input_window_samples=input_window_samples,
final_conv_length="auto",
pool_mode="mean",
second_kernel_size=(2, 32),
third_kernel_size=(8, 4),
drop_prob=0.25)
"""
"""
model = HybridNet(n_chans, n_classes,
input_window_samples=input_window_samples)
"""
"""
model = TCN(n_chans, n_classes,
n_blocks=6,
n_filters=32,
kernel_size=9,
drop_prob=0.0,
add_log_softmax=True)
"""
"""
model = EEGNetv4(n_chans,
n_classes,
input_window_samples=input_window_samples,
final_conv_length="auto",
pool_mode="mean",
F1=8,
D=2,
F2=16, # usually set to F1*D (?)
kernel_length=64,
third_kernel_size=(8, 4),
drop_prob=0.2)
"""
if cuda:
model.cuda()
######################################################################
# Training
# These values we found good for shallow network:
lr = 0.01 # 0.0625 * 0.01
weight_decay = 0.0005
# For deep4 they should be:
# lr = 1 * 0.01
# weight_decay = 0.5 * 0.001
batch_size = 64
n_epochs = 80
clf = EEGClassifier(
model,
criterion=torch.nn.NLLLoss,
optimizer=torch.optim.SGD, #AdamW,
train_split=predefined_split(
valid_set), # using valid_set for validation
optimizer__lr=lr,
optimizer__momentum=0.9,
optimizer__weight_decay=weight_decay,
batch_size=batch_size,
callbacks=[
"accuracy", #("lr_scheduler", LRScheduler('CosineAnnealingLR', T_max=n_epochs - 1)),
],
device=device,
)
# Model training for a specified number of epochs. `y` is None as it is already supplied
# in the dataset.
clf.fit(train_set, y=None, epochs=n_epochs)
results_columns = [
'train_loss', 'valid_loss', 'train_accuracy', 'valid_accuracy'
]
df = pd.DataFrame(clf.history[:, results_columns],
columns=results_columns,
index=clf.history[:, 'epoch'])
val_accs = df['valid_accuracy'].values
max_val_acc = 100.0 * np.max(val_accs)
return max_val_acc
