def create_model_lstm(n_classes, input_time_length, in_chans=22):
set_random_seeds(seed=20170629, cuda=cuda)
# This will determine how many crops are processed in parallel
# final_conv_length determines the size of the receptive field of the ConvNet
lstm_size = 30
lstm_layers = 1
print("#### LSTM SIZE {} ####".format(lstm_size))
print("#### LSTM LAYERS {} ####".format(lstm_layers))
model = ShallowFBCSPLSTM(in_chans=in_chans,
n_classes=n_classes,
input_time_length=input_time_length,
lstm_size=lstm_size,
lstm_layers=lstm_layers,
n_filters_time=lstm_size,
n_filters_spat=lstm_size,
final_conv_length=4,
pool_time_length=20,
pool_time_stride=5).create_network()
to_dense_prediction_model(model)
if cuda:
model.cuda()
return model
def __init__(self, in_chans, n_classes, input_time_length):
super(HybridNetModule, self).__init__()
deep_model = Deep4Net(
in_chans,
n_classes,
n_filters_time=20,
n_filters_spat=30,
n_filters_2=40,
n_filters_3=50,
n_filters_4=60,
input_time_length=input_time_length,
final_conv_length=2,
).create_network()
shallow_model = ShallowFBCSPNet(
in_chans,
n_classes,
input_time_length=input_time_length,
n_filters_time=30,
n_filters_spat=40,
filter_time_length=28,
final_conv_length=29,
).create_network()
reduced_deep_model = nn.Sequential()
for name, module in deep_model.named_children():
if name == "conv_classifier":
new_conv_layer = nn.Conv2d(
module.in_channels,
60,
kernel_size=module.kernel_size,
stride=module.stride,
)
reduced_deep_model.add_module("deep_final_conv",
new_conv_layer)
break
reduced_deep_model.add_module(name, module)
reduced_shallow_model = nn.Sequential()
for name, module in shallow_model.named_children():
if name == "conv_classifier":
new_conv_layer = nn.Conv2d(
module.in_channels,
40,
kernel_size=module.kernel_size,
stride=module.stride,
)
reduced_shallow_model.add_module("shallow_final_conv",
new_conv_layer)
break
reduced_shallow_model.add_module(name, module)
to_dense_prediction_model(reduced_deep_model)
to_dense_prediction_model(reduced_shallow_model)
self.reduced_deep_model = reduced_deep_model
self.reduced_shallow_model = reduced_shallow_model
self.final_conv = nn.Conv2d(100,
n_classes,
kernel_size=(1, 1),
stride=1)
def train(config):
cuda = True
model = config['model']
if model == 'deep':
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length=2,
config=config).create_network()
to_dense_prediction_model(model)
if cuda:
model.cuda()
log.info("Model: \n{:s}".format(str(model)))
dummy_input = np_to_var(train_set.X[:1, :, :, None])
if cuda:
dummy_input = dummy_input.cuda()
out = model(dummy_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
optimizer = optim.Adam(model.parameters())
iterator = CropsFromTrialsIterator(batch_size=60,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
stop_criterion = Or([MaxEpochs(20), NoDecrease('valid_misclass', 80)])
monitors = [
LossMonitor(),
MisclassMonitor(col_suffix='sample_misclass'),
CroppedTrialMisclassMonitor(input_time_length=input_time_length),
RuntimeMonitor()
]
model_constraint = MaxNormDefaultConstraint()
loss_function = lambda preds, targets: F.nll_loss(
th.mean(preds, dim=2, keepdim=False), targets)
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator=iterator,
loss_function=loss_function,
optimizer=optimizer,
model_constraint=model_constraint,
monitors=monitors,
stop_criterion=stop_criterion,
remember_best_column='valid_misclass',
run_after_early_stop=True,
cuda=cuda)
exp.run()
print(exp.rememberer)
return exp.rememberer.lowest_val
def finalized_model_to_dilated(self, model):
to_dense_prediction_model(model)
conv_classifier = model.conv_classifier
model.conv_classifier = nn.Conv2d(conv_classifier.in_channels,
conv_classifier.out_channels,
(global_vars.get('final_conv_size'),
conv_classifier.kernel_size[1]),
stride=conv_classifier.stride,
dilation=conv_classifier.dilation)
def compile(self, loss, optimizer, metrics, cropped=False, seed=0):
self.loss = loss
self._ensure_network_exists()
if cropped:
to_dense_prediction_model(self.network)
if not hasattr(optimizer, 'step'):
optimizer_class = find_optimizer(optimizer)
optimizer = optimizer_class(self.network.parameters())
self.optimizer = optimizer
self.metrics = metrics
self.seed_rng = RandomState(seed)
self.cropped = cropped
self.compiled = True
def compile(
self,
loss,
optimizer,
extra_monitors=None,
cropped=False,
iterator_seed=0,
):
"""
Setup training for this model.
Parameters
----------
loss: function (predictions, targets) -> torch scalar
optimizer: `torch.optim.Optimizer` or string
Either supply an optimizer or the name of the class (e.g. 'adam')
extra_monitors: List of Braindecode monitors, optional
In case you want to monitor additional values except for loss, misclass and runtime.
cropped: bool
Whether to perform cropped decoding, see cropped decoding tutorial.
iterator_seed: int
Seed to seed the iterator random generator.
Returns
-------
"""
self.loss = loss
self._ensure_network_exists()
if cropped:
model_already_dense = np.any(
[
hasattr(m, "dilation")
and (m.dilation != 1)
and (m.dilation) != (1, 1)
for m in self.network.modules()
]
)
if not model_already_dense:
to_dense_prediction_model(self.network)
else:
log.info("Seems model was already converted to dense model...")
if not hasattr(optimizer, "step"):
optimizer_class = find_optimizer(optimizer)
optimizer = optimizer_class(self.network.parameters())
self.optimizer = optimizer
self.extra_monitors = extra_monitors
# Already setting it here, so multiple calls to fit
# will lead to different batches being drawn
self.seed_rng = RandomState(iterator_seed)
self.cropped = cropped
self.compiled = True
def make_regressor(self):
"""
Transforms the model from a network suited for classification into a network suited
for regression by removing the last softmax layer
:return:
"""
if not self.regressed:
new_model = nn.Sequential()
for name, module in self.model.named_children():
# print(name)
if 'softmax' in name:
break
new_model.add_module(name, module)
new_model.add_module('squeeze_out', Expression(squeeze_out))
self.model = new_model
to_dense_prediction_model(self.model)
self.regressed = True
def create_model(n_classes, input_time_length, in_chans=22):
set_random_seeds(seed=20170629, cuda=cuda)
# This will determine how many crops are processed in parallel
# 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=4,
pool_time_length=20,
pool_time_stride=5).create_network()
to_dense_prediction_model(model)
if cuda:
model.cuda()
return model
def build_model(input_time_length, n_chans, n_classes, cropped=False):
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto')
if cropped:
final_conv_length = model.final_conv_length
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length=final_conv_length)
model = model.create_network()
if cropped:
to_dense_prediction_model(model)
return model
def build_cropped_model(model_name, n_chans, n_classes, config):
# input_time_length:
# will determine how many crops are processed in parallel
# supercrop, number of crops taken through network together
# final_conv_length:
# will determine how many crops are processed in parallel
# 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
cropped_input_time_length = config['cropped']['input_time_length']
final_conv_length_shallow = config['cropped']['final_conv_length_shallow']
final_conv_length_deep = config['cropped']['final_conv_length_deep']
if model_name == 'shallow':
model = ShallowFBCSPNet(n_chans, n_classes,
input_time_length=cropped_input_time_length,
final_conv_length=final_conv_length_shallow) \
.create_network()
elif model_name == 'deep':
model = Deep4Net(n_chans, n_classes,
input_time_length=cropped_input_time_length,
final_conv_length=final_conv_length_deep) \
.create_network()
to_dense_prediction_model(model)
return model
def run_exp(max_recording_mins, n_recordings, sec_to_cut,
duration_recording_mins, max_abs_val, max_min_threshold,
max_min_expected, shrink_val, max_min_remove, batch_set_zero_val,
batch_set_zero_test, sampling_freq, low_cut_hz, high_cut_hz,
exp_demean, exp_standardize, moving_demean, moving_standardize,
channel_demean, channel_standardize, divisor, n_folds, i_test_fold,
input_time_length, final_conv_length, pool_stride, n_blocks_to_add,
sigmoid, model_constraint, batch_size, max_epochs,
only_return_exp):
cuda = True
preproc_functions = []
preproc_functions.append(lambda data, fs: (
data[:, int(sec_to_cut * fs):-int(sec_to_cut * fs)], fs))
preproc_functions.append(lambda data, fs: (data[:, :int(
duration_recording_mins * 60 * fs)], fs))
if max_abs_val is not None:
preproc_functions.append(
lambda data, fs: (np.clip(data, -max_abs_val, max_abs_val), fs))
if max_min_threshold is not None:
preproc_functions.append(lambda data, fs: (clean_jumps(
data, 200, max_min_threshold, max_min_expected, cuda), fs))
if max_min_remove is not None:
window_len = 200
preproc_functions.append(lambda data, fs: (set_jumps_to_zero(
data,
window_len=window_len,
threshold=max_min_remove,
cuda=cuda,
clip_min_max_to_zero=True), fs))
if shrink_val is not None:
preproc_functions.append(lambda data, fs: (shrink_spikes(
data,
shrink_val,
1,
9,
), fs))
preproc_functions.append(lambda data, fs: (resampy.resample(
data, fs, sampling_freq, axis=1, filter='kaiser_fast'), sampling_freq))
preproc_functions.append(lambda data, fs: (bandpass_cnt(
data, low_cut_hz, high_cut_hz, fs, filt_order=4, axis=1), fs))
if exp_demean:
preproc_functions.append(lambda data, fs: (exponential_running_demean(
data.T, factor_new=0.001, init_block_size=100).T, fs))
if exp_standardize:
preproc_functions.append(
lambda data, fs: (exponential_running_standardize(
data.T, factor_new=0.001, init_block_size=100).T, fs))
if moving_demean:
preproc_functions.append(lambda data, fs: (padded_moving_demean(
data, axis=1, n_window=201), fs))
if moving_standardize:
preproc_functions.append(lambda data, fs: (padded_moving_standardize(
data, axis=1, n_window=201), fs))
if channel_demean:
preproc_functions.append(lambda data, fs: (demean(data, axis=1), fs))
if channel_standardize:
preproc_functions.append(lambda data, fs:
(standardize(data, axis=1), fs))
if divisor is not None:
preproc_functions.append(lambda data, fs: (data / divisor, fs))
dataset = DiagnosisSet(n_recordings=n_recordings,
max_recording_mins=max_recording_mins,
preproc_functions=preproc_functions)
if not only_return_exp:
X, y = dataset.load()
splitter = Splitter(
n_folds,
i_test_fold,
)
if not only_return_exp:
train_set, valid_set, test_set = splitter.split(X, y)
del X, y # shouldn't be necessary, but just to make sure
else:
train_set = None
valid_set = None
test_set = None
set_random_seeds(seed=20170629, cuda=cuda)
if sigmoid:
n_classes = 1
else:
n_classes = 2
in_chans = 21
net = Deep4Net(
in_chans=in_chans,
n_classes=n_classes,
input_time_length=input_time_length,
final_conv_length=final_conv_length,
pool_time_length=pool_stride,
pool_time_stride=pool_stride,
n_filters_2=50,
n_filters_3=80,
n_filters_4=120,
)
model = net_with_more_layers(net, n_blocks_to_add, nn.MaxPool2d)
if sigmoid:
model = to_linear_plus_minus_net(model)
optimizer = optim.Adam(model.parameters())
to_dense_prediction_model(model)
log.info("Model:\n{:s}".format(str(model)))
if cuda:
model.cuda()
# determine output size
test_input = np_to_var(
np.ones((2, in_chans, input_time_length, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
log.info("{:d} predictions per input/trial".format(n_preds_per_input))
iterator = CropsFromTrialsIterator(batch_size=batch_size,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
if sigmoid:
loss_function = lambda preds, targets: binary_cross_entropy_with_logits(
th.mean(preds, dim=2)[:, 1, 0], targets.type_as(preds))
else:
loss_function = lambda preds, targets: F.nll_loss(
th.mean(preds, dim=2)[:, :, 0], targets)
if model_constraint is not None:
model_constraint = MaxNormDefaultConstraint()
monitors = [
LossMonitor(),
MisclassMonitor(col_suffix='sample_misclass'),
CroppedTrialMisclassMonitor(input_time_length),
RuntimeMonitor(),
]
stop_criterion = MaxEpochs(max_epochs)
batch_modifier = None
if batch_set_zero_val is not None:
batch_modifier = RemoveMinMaxDiff(batch_set_zero_val,
clip_max_abs=True,
set_zero=True)
if (batch_set_zero_val is not None) and (batch_set_zero_test == True):
iterator = ModifiedIterator(
iterator,
batch_modifier,
)
batch_modifier = None
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator,
loss_function,
optimizer,
model_constraint,
monitors,
stop_criterion,
remember_best_column='valid_misclass',
run_after_early_stop=True,
batch_modifier=batch_modifier,
cuda=cuda)
if not only_return_exp:
exp.run()
else:
exp.dataset = dataset
exp.splitter = splitter
return exp
def test_cropped_decoding():
import mne
from mne.io import concatenate_raws
# 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)
# 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=2, feet=3), tmin=1, tmax=4.1,
proj=False, picks=eeg_channel_inds,
baseline=None, preload=True)
import numpy as np
from braindecode.datautil.signal_target import SignalAndTarget
# 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
train_set = SignalAndTarget(X[:60], y=y[:60])
test_set = SignalAndTarget(X[60:], y=y[60:])
from braindecode.models.shallow_fbcsp import ShallowFBCSPNet
from torch import nn
from braindecode.torch_ext.util import set_random_seeds
from braindecode.models.util import to_dense_prediction_model
# 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 = train_set.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).create_network()
to_dense_prediction_model(model)
if cuda:
model.cuda()
from torch import optim
optimizer = optim.Adam(model.parameters())
from braindecode.torch_ext.util import np_to_var
# determine output size
test_input = np_to_var(
np.ones((2, in_chans, input_time_length, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
print("{:d} predictions per input/trial".format(n_preds_per_input))
from braindecode.datautil.iterators import CropsFromTrialsIterator
iterator = CropsFromTrialsIterator(batch_size=32,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
from braindecode.torch_ext.util import np_to_var, var_to_np
import torch.nn.functional as F
from numpy.random import RandomState
import torch as th
from braindecode.experiments.monitors import compute_preds_per_trial_from_crops
rng = RandomState((2017, 6, 30))
losses = []
accuracies = []
for i_epoch in range(4):
# Set model to training mode
model.train()
for batch_X, batch_y in iterator.get_batches(train_set, shuffle=False):
net_in = np_to_var(batch_X)
if cuda:
net_in = net_in.cuda()
net_target = np_to_var(batch_y)
if cuda:
net_target = net_target.cuda()
# Remove gradients of last backward pass from all parameters
optimizer.zero_grad()
outputs = model(net_in)
# Mean predictions across trial
# Note that this will give identical gradients to computing
# a per-prediction loss (at least for the combination of log softmax activation
# and negative log likelihood loss which we are using here)
outputs = th.mean(outputs, dim=2, keepdim=False)
loss = F.nll_loss(outputs, net_target)
loss.backward()
optimizer.step()
# Print some statistics each epoch
model.eval()
print("Epoch {:d}".format(i_epoch))
for setname, dataset in (('Train', train_set), ('Test', test_set)):
# Collect all predictions and losses
all_preds = []
all_losses = []
batch_sizes = []
for batch_X, batch_y in iterator.get_batches(dataset,
shuffle=False):
net_in = np_to_var(batch_X)
if cuda:
net_in = net_in.cuda()
net_target = np_to_var(batch_y)
if cuda:
net_target = net_target.cuda()
outputs = model(net_in)
all_preds.append(var_to_np(outputs))
outputs = th.mean(outputs, dim=2, keepdim=False)
loss = F.nll_loss(outputs, net_target)
loss = float(var_to_np(loss))
all_losses.append(loss)
batch_sizes.append(len(batch_X))
# Compute mean per-input loss
loss = np.mean(np.array(all_losses) * np.array(batch_sizes) /
np.mean(batch_sizes))
print("{:6s} Loss: {:.5f}".format(setname, loss))
losses.append(loss)
# Assign the predictions to the trials
preds_per_trial = compute_preds_per_trial_from_crops(all_preds,
input_time_length,
dataset.X)
# preds per trial are now trials x classes x timesteps/predictions
# Now mean across timesteps for each trial to get per-trial predictions
meaned_preds_per_trial = np.array(
[np.mean(p, axis=1) for p in preds_per_trial])
predicted_labels = np.argmax(meaned_preds_per_trial, axis=1)
accuracy = np.mean(predicted_labels == dataset.y)
accuracies.append(accuracy * 100)
print("{:6s} Accuracy: {:.1f}%".format(
setname, accuracy * 100))
np.testing.assert_allclose(
np.array(losses),
np.array([1.703004002571106,
1.6295261979103088,
0.71168938279151917,
0.70825588703155518,
0.58231228590011597,
0.60176041722297668,
0.46629951894283295,
0.51184913516044617]),
rtol=1e-4, atol=1e-5)
np.testing.assert_allclose(
np.array(accuracies),
np.array(
[50.0,
46.666666666666664,
60.0,
53.333333333333336,
68.333333333333329,
66.666666666666657,
88.333333333333329,
83.333333333333343]),
rtol=1e-4, atol=1e-5)
def test_eeg_classifier():
# 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_window_samples = 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_window_samples=input_window_samples,
final_conv_length=12,
)
to_dense_prediction_model(model)
if cuda:
model.cuda()
# determine output size
test_input = np_to_var(
np.ones((2, in_chans, input_window_samples, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
train_set = create_from_X_y(X[:48],
y[:48],
drop_last_window=False,
window_size_samples=input_window_samples,
window_stride_samples=n_preds_per_input)
valid_set = create_from_X_y(X[48:60],
y[48:60],
drop_last_window=False,
window_size_samples=input_window_samples,
window_stride_samples=n_preds_per_input)
cropped_cb_train = CroppedTrialEpochScoring(
"accuracy",
name="train_trial_accuracy",
lower_is_better=False,
on_train=True,
)
cropped_cb_valid = CroppedTrialEpochScoring(
"accuracy",
on_train=False,
name="valid_trial_accuracy",
lower_is_better=False,
)
clf = EEGClassifier(
model,
cropped=True,
criterion=CroppedLoss,
criterion__loss_function=nll_loss,
optimizer=optim.Adam,
train_split=predefined_split(valid_set),
batch_size=32,
callbacks=[
("train_trial_accuracy", cropped_cb_train),
("valid_trial_accuracy", cropped_cb_valid),
],
)
clf.fit(train_set, y=None, epochs=4)
expected = [{
'batches': [{
'train_batch_size': 32,
'train_loss': 1.6639312505722046
}, {
'train_batch_size': 32,
'train_loss': 2.6161606311798096
}, {
'train_batch_size': 32,
'train_loss': 1.627132773399353
}, {
'valid_batch_size': 24,
'valid_loss': 0.9677614569664001
}],
'epoch':
1,
'train_batch_count':
3,
'train_loss':
1.9690748850504558,
'train_loss_best':
True,
'train_trial_accuracy':
0.4791666666666667,
'train_trial_accuracy_best':
True,
'valid_batch_count':
1,
'valid_loss':
0.9677614569664001,
'valid_loss_best':
True,
'valid_trial_accuracy':
0.5,
'valid_trial_accuracy_best':
True
}, {
'batches': [{
'train_batch_size': 32,
'train_loss': 1.3829222917556763
}, {
'train_batch_size': 32,
'train_loss': 1.3123714923858643
}, {
'train_batch_size': 32,
'train_loss': 1.0109959840774536
}, {
'valid_batch_size': 24,
'valid_loss': 1.9435862302780151
}],
'epoch':
2,
'train_batch_count':
3,
'train_loss':
1.2354299227396648,
'train_loss_best':
True,
'train_trial_accuracy':
0.5,
'train_trial_accuracy_best':
True,
'valid_batch_count':
1,
'valid_loss':
1.9435862302780151,
'valid_loss_best':
False,
'valid_trial_accuracy':
0.5,
'valid_trial_accuracy_best':
False
}, {
'batches': [{
'train_batch_size': 32,
'train_loss': 1.172208547592163
}, {
'train_batch_size': 32,
'train_loss': 0.8899562954902649
}, {
'train_batch_size': 32,
'train_loss': 1.0232216119766235
}, {
'valid_batch_size': 24,
'valid_loss': 0.9585554599761963
}],
'epoch':
3,
'train_batch_count':
3,
'train_loss':
1.0284621516863506,
'train_loss_best':
True,
'train_trial_accuracy':
0.5,
'train_trial_accuracy_best':
False,
'valid_batch_count':
1,
'valid_loss':
0.9585554599761963,
'valid_loss_best':
True,
'valid_trial_accuracy':
0.5,
'valid_trial_accuracy_best':
False
}, {
'batches': [{
'train_batch_size': 32,
'train_loss': 0.9693693518638611
}, {
'train_batch_size': 32,
'train_loss': 0.900641918182373
}, {
'train_batch_size': 32,
'train_loss': 0.8839665651321411
}, {
'valid_batch_size': 24,
'valid_loss': 0.873468816280365
}],
'epoch':
4,
'train_batch_count':
3,
'train_loss':
0.9179926117261251,
'train_loss_best':
True,
'train_trial_accuracy':
0.625,
'train_trial_accuracy_best':
True,
'valid_batch_count':
1,
'valid_loss':
0.873468816280365,
'valid_loss_best':
True,
'valid_trial_accuracy':
0.4166666666666667,
'valid_trial_accuracy_best':
False
}]
history_without_dur = [{k: v
for k, v in h.items() if k != "dur"}
for h in clf.history]
assert_deep_allclose(expected, history_without_dur, atol=1e-3, rtol=1e-3)
def runModel(mode):
cudnn.benchmark = True
start = time.time()
#mode = str(sys.argv[1])
#X,y,test_X,test_y = loadSubNormData(mode='all')
#X,y,test_X,test_y = loadNEDCdata(mode=mode)
#data = np.load('sessionsData/data%s-sessions.npy'%mode[:3])
#labels = np.load('sessionsData/labels%s-sessions.npy'%mode[:3])
data = np.load('data%s.npy' % mode[:3])
labels = np.load('labels%s.npy' % mode[:3])
X, y, test_X, test_y = splitDataRandom_Loaded(data, labels, mode)
print('Mode - %s Total n: %d, Test n: %d' %
(mode, len(y) + len(test_y), len(test_y)))
#return 0
#X = addDataNoise(X,band=[1,4])
#test_X = addDataNoise(test_X,band=[1,4])
max_shape = np.max([list(x.shape) for x in X], axis=0)
assert max_shape[1] == int(config.duration_recording_mins *
config.sampling_freq * 60)
n_classes = 2
n_recordings = None # set to an integer, if you want to restrict the set size
sensor_types = ["EEG"]
n_chans = 19 #21
max_recording_mins = 35 # exclude larger recordings from training set
sec_to_cut = 60 # cut away at start of each recording
duration_recording_mins = 5 #20 # how many minutes to use per recording
test_recording_mins = 5 #20
max_abs_val = 800 # for clipping
sampling_freq = 100
divisor = 10 # divide signal by this
test_on_eval = True # teston evaluation set or on training set
# in case of test on eval, n_folds and i_testfold determine
# validation fold in training set for training until first stop
n_folds = 10
i_test_fold = 9
shuffle = True
model_name = 'linear' #'deep'#'shallow' 'linear'
n_start_chans = 25
n_chan_factor = 2 # relevant for deep model only
input_time_length = 6000
final_conv_length = 1
model_constraint = 'defaultnorm'
init_lr = 1e-3
batch_size = 64
max_epochs = 35 # until first stop, the continue train on train+valid
cuda = True # False
if model_name == 'shallow':
model = ShallowFBCSPNet(
in_chans=n_chans,
n_classes=n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
final_conv_length=final_conv_length).create_network()
elif model_name == 'deep':
model = Deep4Net(n_chans,
n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
n_filters_2=int(n_start_chans * n_chan_factor),
n_filters_3=int(n_start_chans * (n_chan_factor**2.0)),
n_filters_4=int(n_start_chans * (n_chan_factor**3.0)),
final_conv_length=final_conv_length,
stride_before_pool=True).create_network()
elif (model_name == 'deep_smac'):
if model_name == 'deep_smac':
do_batch_norm = False
else:
assert model_name == 'deep_smac_bnorm'
do_batch_norm = True
double_time_convs = False
drop_prob = 0.244445
filter_length_2 = 12
filter_length_3 = 14
filter_length_4 = 12
filter_time_length = 21
final_conv_length = 1
first_nonlin = elu
first_pool_mode = 'mean'
first_pool_nonlin = identity
later_nonlin = elu
later_pool_mode = 'mean'
later_pool_nonlin = identity
n_filters_factor = 1.679066
n_filters_start = 32
pool_time_length = 1
pool_time_stride = 2
split_first_layer = True
n_chan_factor = n_filters_factor
n_start_chans = n_filters_start
model = Deep4Net(n_chans,
n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
n_filters_2=int(n_start_chans * n_chan_factor),
n_filters_3=int(n_start_chans * (n_chan_factor**2.0)),
n_filters_4=int(n_start_chans * (n_chan_factor**3.0)),
final_conv_length=final_conv_length,
batch_norm=do_batch_norm,
double_time_convs=double_time_convs,
drop_prob=drop_prob,
filter_length_2=filter_length_2,
filter_length_3=filter_length_3,
filter_length_4=filter_length_4,
filter_time_length=filter_time_length,
first_nonlin=first_nonlin,
first_pool_mode=first_pool_mode,
first_pool_nonlin=first_pool_nonlin,
later_nonlin=later_nonlin,
later_pool_mode=later_pool_mode,
later_pool_nonlin=later_pool_nonlin,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
split_first_layer=split_first_layer,
stride_before_pool=True).create_network()
elif model_name == 'shallow_smac':
conv_nonlin = identity
do_batch_norm = True
drop_prob = 0.328794
filter_time_length = 56
final_conv_length = 22
n_filters_spat = 73
n_filters_time = 24
pool_mode = 'max'
pool_nonlin = identity
pool_time_length = 84
pool_time_stride = 3
split_first_layer = True
model = ShallowFBCSPNet(
in_chans=n_chans,
n_classes=n_classes,
n_filters_time=n_filters_time,
n_filters_spat=n_filters_spat,
input_time_length=input_time_length,
final_conv_length=final_conv_length,
conv_nonlin=conv_nonlin,
batch_norm=do_batch_norm,
drop_prob=drop_prob,
filter_time_length=filter_time_length,
pool_mode=pool_mode,
pool_nonlin=pool_nonlin,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
split_first_layer=split_first_layer,
).create_network()
elif model_name == 'linear':
model = nn.Sequential()
model.add_module("conv_classifier",
nn.Conv2d(n_chans, n_classes, (600, 1)))
model.add_module('softmax', nn.LogSoftmax(dim=1))
model.add_module('squeeze', Expression(lambda x: x.squeeze(3)))
else:
assert False, "unknown model name {:s}".format(model_name)
to_dense_prediction_model(model)
if config.cuda:
model.cuda()
test_input = np_to_var(
np.ones((2, config.n_chans, config.input_time_length, 1),
dtype=np.float32))
if config.cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
iterator = CropsFromTrialsIterator(
batch_size=config.batch_size,
input_time_length=config.input_time_length,
n_preds_per_input=n_preds_per_input)
#model.add_module('softmax', nn.LogSoftmax(dim=1))
model.eval()
mode[2] = str(mode[2])
mode[3] = str(mode[3])
modelName = '-'.join(mode[:4])
#params = th.load('sessionsData/%sModel%s-sessions.pt'%(modelName,mode[4]))
#params = th.load('%sModel%s.pt'%(modelName,mode[4]))
params = th.load('linear/%sModel%s.pt' % (modelName, mode[4]))
model.load_state_dict(params)
if config.test_on_eval:
#test_X, test_y = test_dataset.load()
#test_X, test_y = loadNEDCdata(mode='eval')
max_shape = np.max([list(x.shape) for x in test_X], axis=0)
assert max_shape[1] == int(config.test_recording_mins *
config.sampling_freq * 60)
if not config.test_on_eval:
splitter = TrainValidTestSplitter(config.n_folds,
config.i_test_fold,
shuffle=config.shuffle)
train_set, valid_set, test_set = splitter.split(X, y)
else:
splitter = TrainValidSplitter(config.n_folds,
i_valid_fold=config.i_test_fold,
shuffle=config.shuffle)
train_set, valid_set = splitter.split(X, y)
test_set = SignalAndTarget(test_X, test_y)
del test_X, test_y
del X, y # shouldn't be necessary, but just to make sure
datasets = OrderedDict(
(('train', train_set), ('valid', valid_set), ('test', test_set)))
for setname in ('train', 'valid', 'test'):
#setname = 'test'
#print("Compute predictions for {:s}...".format(setname))
dataset = datasets[setname]
if config.cuda:
preds_per_batch = [
var_to_np(model(np_to_var(b[0]).cuda()))
for b in iterator.get_batches(dataset, shuffle=False)
]
else:
preds_per_batch = [
var_to_np(model(np_to_var(b[0])))
for b in iterator.get_batches(dataset, shuffle=False)
]
preds_per_trial = compute_preds_per_trial(
preds_per_batch,
dataset,
input_time_length=iterator.input_time_length,
n_stride=iterator.n_preds_per_input)
mean_preds_per_trial = [
np.mean(preds, axis=(0, 2)) for preds in preds_per_trial
]
mean_preds_per_trial = np.array(mean_preds_per_trial)
all_pred_labels = np.argmax(mean_preds_per_trial, axis=1).squeeze()
all_target_labels = dataset.y
acc_per_class = []
for i_class in range(n_classes):
mask = all_target_labels == i_class
acc = np.mean(all_pred_labels[mask] == all_target_labels[mask])
acc_per_class.append(acc)
misclass = 1 - np.mean(acc_per_class)
#print('Acc:{}, Class 0:{}, Class 1:{}'.format(np.mean(acc_per_class),acc_per_class[0],acc_per_class[1]))
if setname == 'test':
testResult = np.mean(acc_per_class)
return testResult
def exp(subject_id):
import torch
test_subj = np.r_[subject_id]
print('test subj:' + str(test_subj))
#20% validation
train_size = int(0.9* len(splitted['session_T']))
test_size = len(splitted['session_T']) - train_size
# train_set, valid_set = torch.utils.data.random_split(splitted['session_T'], [train_size, test_size])
train_set = splitted['session_T']
test_set = splitted['session_E']
# model = Deep4Net(
# n_chans,
# n_classes,
# input_window_samples=input_window_samples,
# final_conv_length="auto",
# )
from torch.utils.data import Dataset, ConcatDataset
crop_size = 1000
# embedding_net = Deep4Net_origin(n_classes, n_chans, crop_size)
# model = FcClfNet(embedding_net)
model = ShallowFBCSPNet(
n_chans,
n_classes,
input_window_samples=input_window_samples,
final_conv_length='auto',
)
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, 22, input_window_samples)[2]
print("n_preds_per_input : ", n_preds_per_input)
print(model)
batch_size =8
epochs = 200
lr = 0.0625 * 0.01
weight_decay = 0
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True)
# valid_loader = torch.utils.data.DataLoader(valid_set, batch_size=batch_size, shuffle=False)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=False)
# Send model to GPU
if cuda:
model.cuda()
from torch.optim import lr_scheduler
import torch.optim as optim
import argparse
parser = argparse.ArgumentParser(description='cross subject domain adaptation')
parser.add_argument('--batch-size', type=int, default=50, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=50, metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=100, metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr', type=float, default=0.001, metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model', action='store_true', default=True,
help='For Saving the current Model')
args = parser.parse_args()
args.gpuidx = 0
args.seed = 0
args.use_tensorboard = False
args.save_model = False
optimizer = optim.AdamW(model.parameters(), lr=0.01, weight_decay=0.5 * 0.001)
# scheduler = lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10, T_mult=1)
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=epochs-1)
import pandas as pd
results_columns = ['test_loss', 'test_accuracy']
df = pd.DataFrame(columns=results_columns)
for epochidx in range(1, epochs):
print(epochidx)
train_crop(10, model, device, train_loader,optimizer,scheduler,cuda, args.gpuidx)
test_loss, test_score = eval_crop(model, device, test_loader)
results = { 'test_loss': test_loss, 'test_accuracy': test_score}
df = df.append(results, ignore_index=True)
print(results)
return df
def run_exp(max_recording_mins, n_recordings, sec_to_cut,
duration_recording_mins, max_abs_val, shrink_val, sampling_freq,
divisor, n_folds, i_test_fold, final_conv_length, model_constraint,
batch_size, max_epochs, n_filters_time, n_filters_spat,
filter_time_length, conv_nonlin, pool_time_length,
pool_time_stride, pool_mode, pool_nonlin, split_first_layer,
do_batch_norm, drop_prob, time_cut_off_sec, start_time,
input_time_length, only_return_exp):
kwargs = locals()
for model_param in [
'final_conv_length',
'n_filters_time',
'n_filters_spat',
'filter_time_length',
'conv_nonlin',
'pool_time_length',
'pool_time_stride',
'pool_mode',
'pool_nonlin',
'split_first_layer',
'do_batch_norm',
'drop_prob',
]:
kwargs.pop(model_param)
nonlin_dict = {
'elu': elu,
'relu': relu,
'relu6': relu6,
'tanh': tanh,
'square': square,
'identity': identity,
'log': safe_log,
}
assert input_time_length == 6000
# copy over from early seizure
# make proper
n_classes = 2
in_chans = 21
cuda = True
set_random_seeds(seed=20170629, cuda=cuda)
model = ShallowFBCSPNet(in_chans=in_chans,
n_classes=n_classes,
input_time_length=input_time_length,
final_conv_length=final_conv_length,
n_filters_time=n_filters_time,
filter_time_length=filter_time_length,
n_filters_spat=n_filters_spat,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
conv_nonlin=nonlin_dict[conv_nonlin],
pool_mode=pool_mode,
pool_nonlin=nonlin_dict[pool_nonlin],
split_first_layer=split_first_layer,
batch_norm=do_batch_norm,
batch_norm_alpha=0.1,
drop_prob=drop_prob).create_network()
to_dense_prediction_model(model)
if cuda:
model.cuda()
model.eval()
test_input = np_to_var(
np.ones((2, in_chans, input_time_length, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
try:
out = model(test_input)
except RuntimeError:
raise ValueError("Model receptive field too large...")
n_preds_per_input = out.cpu().data.numpy().shape[2]
n_receptive_field = input_time_length - n_preds_per_input
if n_receptive_field > 6000:
raise ValueError("Model receptive field ({:d}) too large...".format(
n_receptive_field))
# For future, here optionally add input time length instead
model = ShallowFBCSPNet(in_chans=in_chans,
n_classes=n_classes,
input_time_length=input_time_length,
final_conv_length=final_conv_length,
n_filters_time=n_filters_time,
filter_time_length=filter_time_length,
n_filters_spat=n_filters_spat,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
conv_nonlin=nonlin_dict[conv_nonlin],
pool_mode=pool_mode,
pool_nonlin=nonlin_dict[pool_nonlin],
split_first_layer=split_first_layer,
batch_norm=do_batch_norm,
batch_norm_alpha=0.1,
drop_prob=drop_prob).create_network()
return common.run_exp(model=model, **kwargs)
def run_exp(
data_folders,
n_recordings,
sensor_types,
n_chans,
max_recording_mins,
sec_to_cut,
duration_recording_mins,
test_recording_mins,
max_abs_val,
sampling_freq,
divisor,
test_on_eval,
n_folds,
i_test_fold,
shuffle,
model_name,
n_start_chans,
n_chan_factor,
input_time_length,
final_conv_length,
model_constraint,
init_lr,
batch_size,
max_epochs,
cuda,
):
import torch.backends.cudnn as cudnn
cudnn.benchmark = True
preproc_functions = []
preproc_functions.append(lambda data, fs: (
data[:, int(sec_to_cut * fs):-int(sec_to_cut * fs)], fs))
preproc_functions.append(lambda data, fs: (data[:, :int(
duration_recording_mins * 60 * fs)], fs))
if max_abs_val is not None:
preproc_functions.append(
lambda data, fs: (np.clip(data, -max_abs_val, max_abs_val), fs))
preproc_functions.append(lambda data, fs: (resampy.resample(
data, fs, sampling_freq, axis=1, filter='kaiser_fast'), sampling_freq))
if divisor is not None:
preproc_functions.append(lambda data, fs: (data / divisor, fs))
dataset = DiagnosisSet(n_recordings=n_recordings,
max_recording_mins=max_recording_mins,
preproc_functions=preproc_functions,
data_folders=data_folders,
train_or_eval='train',
sensor_types=sensor_types)
if test_on_eval:
if test_recording_mins is None:
test_recording_mins = duration_recording_mins
test_preproc_functions = copy(preproc_functions)
test_preproc_functions[1] = lambda data, fs: (data[:, :int(
test_recording_mins * 60 * fs)], fs)
test_dataset = DiagnosisSet(n_recordings=n_recordings,
max_recording_mins=None,
preproc_functions=test_preproc_functions,
data_folders=data_folders,
train_or_eval='eval',
sensor_types=sensor_types)
X, y = dataset.load()
max_shape = np.max([list(x.shape) for x in X], axis=0)
assert max_shape[1] == int(duration_recording_mins * sampling_freq * 60)
if test_on_eval:
test_X, test_y = test_dataset.load()
max_shape = np.max([list(x.shape) for x in test_X], axis=0)
assert max_shape[1] == int(test_recording_mins * sampling_freq * 60)
if not test_on_eval:
splitter = TrainValidTestSplitter(n_folds,
i_test_fold,
shuffle=shuffle)
train_set, valid_set, test_set = splitter.split(X, y)
else:
splitter = TrainValidSplitter(n_folds,
i_valid_fold=i_test_fold,
shuffle=shuffle)
train_set, valid_set = splitter.split(X, y)
test_set = SignalAndTarget(test_X, test_y)
del test_X, test_y
del X, y # shouldn't be necessary, but just to make sure
set_random_seeds(seed=20170629, cuda=cuda)
n_classes = 2
if model_name == 'shallow':
model = ShallowFBCSPNet(
in_chans=n_chans,
n_classes=n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
final_conv_length=final_conv_length).create_network()
elif model_name == 'deep':
model = Deep4Net(n_chans,
n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
n_filters_2=int(n_start_chans * n_chan_factor),
n_filters_3=int(n_start_chans * (n_chan_factor**2.0)),
n_filters_4=int(n_start_chans * (n_chan_factor**3.0)),
final_conv_length=final_conv_length,
stride_before_pool=True).create_network()
elif (model_name == 'deep_smac'):
if model_name == 'deep_smac':
do_batch_norm = False
else:
assert model_name == 'deep_smac_bnorm'
do_batch_norm = True
double_time_convs = False
drop_prob = 0.244445
filter_length_2 = 12
filter_length_3 = 14
filter_length_4 = 12
filter_time_length = 21
final_conv_length = 1
first_nonlin = elu
first_pool_mode = 'mean'
first_pool_nonlin = identity
later_nonlin = elu
later_pool_mode = 'mean'
later_pool_nonlin = identity
n_filters_factor = 1.679066
n_filters_start = 32
pool_time_length = 1
pool_time_stride = 2
split_first_layer = True
n_chan_factor = n_filters_factor
n_start_chans = n_filters_start
model = Deep4Net(n_chans,
n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
n_filters_2=int(n_start_chans * n_chan_factor),
n_filters_3=int(n_start_chans * (n_chan_factor**2.0)),
n_filters_4=int(n_start_chans * (n_chan_factor**3.0)),
final_conv_length=final_conv_length,
batch_norm=do_batch_norm,
double_time_convs=double_time_convs,
drop_prob=drop_prob,
filter_length_2=filter_length_2,
filter_length_3=filter_length_3,
filter_length_4=filter_length_4,
filter_time_length=filter_time_length,
first_nonlin=first_nonlin,
first_pool_mode=first_pool_mode,
first_pool_nonlin=first_pool_nonlin,
later_nonlin=later_nonlin,
later_pool_mode=later_pool_mode,
later_pool_nonlin=later_pool_nonlin,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
split_first_layer=split_first_layer,
stride_before_pool=True).create_network()
elif model_name == 'shallow_smac':
conv_nonlin = identity
do_batch_norm = True
drop_prob = 0.328794
filter_time_length = 56
final_conv_length = 22
n_filters_spat = 73
n_filters_time = 24
pool_mode = 'max'
pool_nonlin = identity
pool_time_length = 84
pool_time_stride = 3
split_first_layer = True
model = ShallowFBCSPNet(
in_chans=n_chans,
n_classes=n_classes,
n_filters_time=n_filters_time,
n_filters_spat=n_filters_spat,
input_time_length=input_time_length,
final_conv_length=final_conv_length,
conv_nonlin=conv_nonlin,
batch_norm=do_batch_norm,
drop_prob=drop_prob,
filter_time_length=filter_time_length,
pool_mode=pool_mode,
pool_nonlin=pool_nonlin,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
split_first_layer=split_first_layer,
).create_network()
elif model_name == 'linear':
model = nn.Sequential()
model.add_module("conv_classifier",
nn.Conv2d(n_chans, n_classes, (600, 1)))
model.add_module('softmax', nn.LogSoftmax())
model.add_module('squeeze', Expression(lambda x: x.squeeze(3)))
else:
assert False, "unknown model name {:s}".format(model_name)
to_dense_prediction_model(model)
log.info("Model:\n{:s}".format(str(model)))
if cuda:
model.cuda()
# determine output size
test_input = np_to_var(
np.ones((2, n_chans, input_time_length, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
log.info("In shape: {:s}".format(str(test_input.cpu().data.numpy().shape)))
out = model(test_input)
log.info("Out shape: {:s}".format(str(out.cpu().data.numpy().shape)))
n_preds_per_input = out.cpu().data.numpy().shape[2]
log.info("{:d} predictions per input/trial".format(n_preds_per_input))
iterator = CropsFromTrialsIterator(batch_size=batch_size,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
optimizer = optim.Adam(model.parameters(), lr=init_lr)
loss_function = lambda preds, targets: F.nll_loss(
th.mean(preds, dim=2, keepdim=False), targets)
if model_constraint is not None:
assert model_constraint == 'defaultnorm'
model_constraint = MaxNormDefaultConstraint()
monitors = [
LossMonitor(),
MisclassMonitor(col_suffix='sample_misclass'),
CroppedDiagnosisMonitor(input_time_length, n_preds_per_input),
RuntimeMonitor(),
]
stop_criterion = MaxEpochs(max_epochs)
batch_modifier = None
run_after_early_stop = True
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator,
loss_function,
optimizer,
model_constraint,
monitors,
stop_criterion,
remember_best_column='valid_misclass',
run_after_early_stop=run_after_early_stop,
batch_modifier=batch_modifier,
cuda=cuda)
exp.run()
return exp
def run_exp(test_on_eval, sensor_types, n_chans, max_recording_mins,
test_recording_mins, n_recordings, sec_to_cut_at_start,
sec_to_cut_at_end, duration_recording_mins, max_abs_val,
clip_before_resample, sampling_freq, divisor, n_folds, i_test_fold,
shuffle, merge_train_valid, model_name, n_start_chans,
n_chan_factor, input_time_length, final_conv_length,
stride_before_pool, optimizer, learning_rate, weight_decay,
scheduler, model_constraint, batch_size, max_epochs, log_dir,
only_return_exp, np_th_seed):
cuda = True
if ('smac' in model_name) and (input_time_length is None):
input_time_length = 12000
fix_input_length_for_smac = True
else:
fix_input_length_for_smac = False
set_random_seeds(seed=np_th_seed, cuda=cuda)
n_classes = 2
if model_name == 'shallow':
model = ShallowFBCSPNet(
in_chans=n_chans,
n_classes=n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
final_conv_length=final_conv_length).create_network()
elif model_name == 'deep':
model = Deep4Net(
n_chans,
n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
n_filters_2=int(n_start_chans * n_chan_factor),
n_filters_3=int(n_start_chans * (n_chan_factor**2.0)),
n_filters_4=int(n_start_chans * (n_chan_factor**3.0)),
final_conv_length=final_conv_length,
stride_before_pool=stride_before_pool).create_network()
elif (model_name == 'deep_smac') or (model_name == 'deep_smac_bnorm'):
if model_name == 'deep_smac':
do_batch_norm = False
else:
assert model_name == 'deep_smac_bnorm'
do_batch_norm = True
double_time_convs = False
drop_prob = 0.244445
filter_length_2 = 12
filter_length_3 = 14
filter_length_4 = 12
filter_time_length = 21
final_conv_length = 1
first_nonlin = elu
first_pool_mode = 'mean'
first_pool_nonlin = identity
later_nonlin = elu
later_pool_mode = 'mean'
later_pool_nonlin = identity
n_filters_factor = 1.679066
n_filters_start = 32
pool_time_length = 1
pool_time_stride = 2
split_first_layer = True
n_chan_factor = n_filters_factor
n_start_chans = n_filters_start
model = Deep4Net(n_chans,
n_classes,
n_filters_time=n_start_chans,
n_filters_spat=n_start_chans,
input_time_length=input_time_length,
n_filters_2=int(n_start_chans * n_chan_factor),
n_filters_3=int(n_start_chans * (n_chan_factor**2.0)),
n_filters_4=int(n_start_chans * (n_chan_factor**3.0)),
final_conv_length=final_conv_length,
batch_norm=do_batch_norm,
double_time_convs=double_time_convs,
drop_prob=drop_prob,
filter_length_2=filter_length_2,
filter_length_3=filter_length_3,
filter_length_4=filter_length_4,
filter_time_length=filter_time_length,
first_nonlin=first_nonlin,
first_pool_mode=first_pool_mode,
first_pool_nonlin=first_pool_nonlin,
later_nonlin=later_nonlin,
later_pool_mode=later_pool_mode,
later_pool_nonlin=later_pool_nonlin,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
split_first_layer=split_first_layer,
stride_before_pool=True).create_network()
elif model_name == 'shallow_smac':
conv_nonlin = identity
do_batch_norm = True
drop_prob = 0.328794
filter_time_length = 56
final_conv_length = 22
n_filters_spat = 73
n_filters_time = 24
pool_mode = 'max'
pool_nonlin = identity
pool_time_length = 84
pool_time_stride = 3
split_first_layer = True
model = ShallowFBCSPNet(
in_chans=n_chans,
n_classes=n_classes,
n_filters_time=n_filters_time,
n_filters_spat=n_filters_spat,
input_time_length=input_time_length,
final_conv_length=final_conv_length,
conv_nonlin=conv_nonlin,
batch_norm=do_batch_norm,
drop_prob=drop_prob,
filter_time_length=filter_time_length,
pool_mode=pool_mode,
pool_nonlin=pool_nonlin,
pool_time_length=pool_time_length,
pool_time_stride=pool_time_stride,
split_first_layer=split_first_layer,
).create_network()
elif model_name == 'deep_smac_new':
from torch.nn.functional import elu, relu, relu6, tanh
from braindecode.torch_ext.functions import identity, square, safe_log
n_filters_factor = 1.9532637176784269
n_filters_start = 61
deep_kwargs = {
"batch_norm": False,
"double_time_convs": False,
"drop_prob": 0.3622676569047184,
"filter_length_2": 9,
"filter_length_3": 6,
"filter_length_4": 10,
"filter_time_length": 17,
"final_conv_length": 5,
"first_nonlin": elu,
"first_pool_mode": "max",
"first_pool_nonlin": identity,
"later_nonlin": relu6,
"later_pool_mode": "max",
"later_pool_nonlin": identity,
"n_filters_time": n_filters_start,
"n_filters_spat": n_filters_start,
"n_filters_2": int(n_filters_start * n_filters_factor),
"n_filters_3": int(n_filters_start * (n_filters_factor**2.0)),
"n_filters_4": int(n_filters_start * (n_filters_factor**3.0)),
"pool_time_length": 1,
"pool_time_stride": 4,
"split_first_layer": True,
"stride_before_pool": True,
}
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
**deep_kwargs).create_network()
elif model_name == 'shallow_smac_new':
from torch.nn.functional import elu, relu, relu6, tanh
from braindecode.torch_ext.functions import identity, square, safe_log
shallow_kwargs = {
"conv_nonlin": square,
"batch_norm": True,
"drop_prob": 0.10198630723385381,
"filter_time_length": 51,
"final_conv_length": 1,
"n_filters_spat": 200,
"n_filters_time": 76,
"pool_mode": "max",
"pool_nonlin": safe_log,
"pool_time_length": 139,
"pool_time_stride": 49,
"split_first_layer": True,
}
model = ShallowFBCSPNet(in_chans=n_chans,
n_classes=n_classes,
input_time_length=input_time_length,
**shallow_kwargs).create_network()
elif model_name == 'linear':
model = nn.Sequential()
model.add_module("conv_classifier",
nn.Conv2d(n_chans, n_classes, (600, 1)))
model.add_module('softmax', nn.LogSoftmax())
model.add_module('squeeze', Expression(lambda x: x.squeeze(3)))
elif model_name == '3path':
virtual_chan_1x1_conv = True
mean_across_features = False
drop_prob = 0.5
n_start_filters = 10
early_bnorm = False
n_classifier_filters = 100
later_kernel_len = 5
extra_conv_stride = 4
# dont forget to reset n_preds_per_blabla
model = create_multi_start_path_net(
in_chans=n_chans,
virtual_chan_1x1_conv=virtual_chan_1x1_conv,
n_start_filters=n_start_filters,
early_bnorm=early_bnorm,
later_kernel_len=later_kernel_len,
extra_conv_stride=extra_conv_stride,
mean_across_features=mean_across_features,
n_classifier_filters=n_classifier_filters,
drop_prob=drop_prob)
else:
assert False, "unknown model name {:s}".format(model_name)
if not model_name == '3path':
to_dense_prediction_model(model)
log.info("Model:\n{:s}".format(str(model)))
time_cut_off_sec = np.inf
start_time = time.time()
# fix input time length in case of smac models
if fix_input_length_for_smac:
assert ('smac' in model_name) and (input_time_length == 12000)
if cuda:
model.cuda()
test_input = np_to_var(
np.ones((2, n_chans, input_time_length, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
try:
out = model(test_input)
except:
raise ValueError("Model receptive field too large...")
n_preds_per_input = out.cpu().data.numpy().shape[2]
n_receptive_field = input_time_length - n_preds_per_input
input_time_length = 2 * n_receptive_field
exp = common.run_exp(
max_recording_mins,
n_recordings,
sec_to_cut_at_start,
sec_to_cut_at_end,
duration_recording_mins,
max_abs_val,
clip_before_resample,
sampling_freq,
divisor,
n_folds,
i_test_fold,
shuffle,
merge_train_valid,
model,
input_time_length,
optimizer,
learning_rate,
weight_decay,
scheduler,
model_constraint,
batch_size,
max_epochs,
only_return_exp,
time_cut_off_sec,
start_time,
test_on_eval,
test_recording_mins,
sensor_types,
log_dir,
np_th_seed,
)
return exp
def test_predict_trials():
ds = MOABBDataset('BNCI2014001', subject_ids=1)
ds1 = ds.split([0])['0']
# determine original trial size
windows_ds1 = create_windows_from_events(
ds1,
)
trial_size = windows_ds1[0][0].shape[1]
# create two windows per trial, where windows maximally overlap
window_size_samples = trial_size - 1
window_stride_samples = 5
windows_ds1 = create_windows_from_events(
ds1,
window_size_samples=window_size_samples,
window_stride_samples=window_stride_samples,
drop_last_window=False,
)
in_chans = windows_ds1[0][0].shape[0]
n_classes = len(windows_ds1.get_metadata()['target'].unique())
model = ShallowFBCSPNet(
in_chans=in_chans,
n_classes=n_classes,
)
to_dense_prediction_model(model)
output_shape = get_output_shape(model, in_chans, window_size_samples)
# the number of samples required to get 1 output
receptive_field_size = window_size_samples - output_shape[-1] + 1
preds, targets = predict_trials(model, windows_ds1)
# some model, cropped data
assert preds.shape[-1] + receptive_field_size - 1 == trial_size
assert preds.shape[1] == n_classes
assert preds.shape[0] == targets.shape[0]
metadata = windows_ds1.get_metadata()
expected_targets = metadata[metadata['i_window_in_trial'] == 0][
'target'].values
np.testing.assert_array_equal(expected_targets, targets)
# some model, trialwise data
windows_ds2 = create_windows_from_events(ds1)
with pytest.warns(UserWarning, match='This function was designed to predict'
' trials from cropped datasets.'):
predict_trials(model, windows_ds2)
# cropped EEGClassifier, cropped data
clf = EEGClassifier(
model,
criterion=torch.nn.NLLLoss,
optimizer=optim.AdamW,
train_split=None,
optimizer__lr=0.0625 * 0.01,
optimizer__weight_decay=0,
batch_size=64,
)
clf.initialize()
clf.predict_trials(windows_ds1, return_targets=True)
# cropped EEGClassifier, trialwise data
with pytest.warns(UserWarning, match="This method was designed to predict "
"trials in cropped mode. Calling it "
"when cropped is False will give the "
"same result as '.predict'."):
clf.predict_trials(windows_ds2)
def test_eeg_classifier():
# 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()
# determine output size
test_input = np_to_var(
np.ones((2, in_chans, input_time_length, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
train_set = CroppedXyDataset(X[:60],
y=y[:60],
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
cropped_cb_train = CroppedTrialEpochScoring(
"accuracy",
name="train_trial_accuracy",
lower_is_better=False,
on_train=True,
)
cropped_cb_valid = CroppedTrialEpochScoring(
"accuracy",
on_train=False,
name="valid_trial_accuracy",
lower_is_better=False,
)
clf = EEGClassifier(
model,
criterion=CroppedNLLLoss,
optimizer=optim.Adam,
train_split=TrainTestSplit(
train_size=0.8,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input,
),
batch_size=32,
callbacks=[
("train_trial_accuracy", cropped_cb_train),
("valid_trial_accuracy", cropped_cb_valid),
],
)
clf.fit(train_set.X, train_set.y, epochs=4)
expected = [
{
"batches": [
{
"train_loss": 2.0750882625579834,
"train_batch_size": 32
},
{
"train_loss": 3.09424090385437,
"train_batch_size": 32
},
{
"train_loss": 1.079931378364563,
"train_batch_size": 32
},
{
"valid_loss": 2.3208131790161133,
"valid_batch_size": 24
},
],
"epoch":
1,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
2.083086848258972,
"train_loss_best":
True,
"valid_loss":
2.3208131790161133,
"valid_loss_best":
True,
"train_trial_accuracy":
0.5,
"train_trial_accuracy_best":
True,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
True,
},
{
"batches": [
{
"train_loss": 1.827332615852356,
"train_batch_size": 32
},
{
"train_loss": 1.4135494232177734,
"train_batch_size": 32
},
{
"train_loss": 1.1295170783996582,
"train_batch_size": 32
},
{
"valid_loss": 1.4291356801986694,
"valid_batch_size": 24
},
],
"epoch":
2,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
1.4567997058232625,
"train_loss_best":
True,
"valid_loss":
1.4291356801986694,
"valid_loss_best":
True,
"train_trial_accuracy":
0.5,
"train_trial_accuracy_best":
False,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
False,
},
{
"batches": [
{
"train_loss": 1.1495535373687744,
"train_batch_size": 32
},
{
"train_loss": 2.356320381164551,
"train_batch_size": 32
},
{
"train_loss": 0.9548418521881104,
"train_batch_size": 32
},
{
"valid_loss": 2.248246908187866,
"valid_batch_size": 24
},
],
"epoch":
3,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
1.4869052569071453,
"train_loss_best":
False,
"valid_loss":
2.248246908187866,
"valid_loss_best":
False,
"train_trial_accuracy":
0.5,
"train_trial_accuracy_best":
False,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
False,
},
{
"batches": [
{
"train_loss": 1.2157528400421143,
"train_batch_size": 32
},
{
"train_loss": 1.1182057857513428,
"train_batch_size": 32
},
{
"train_loss": 0.9163083434104919,
"train_batch_size": 32
},
{
"valid_loss": 0.9732739925384521,
"valid_batch_size": 24
},
],
"epoch":
4,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
1.083422323067983,
"train_loss_best":
True,
"valid_loss":
0.9732739925384521,
"valid_loss_best":
True,
"train_trial_accuracy":
0.5,
"train_trial_accuracy_best":
False,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
False,
},
]
history_without_dur = [{k: v
for k, v in h.items() if k != "dur"}
for h in clf.history]
assert_deep_allclose(history_without_dur, expected, atol=1e-3, rtol=1e-3)
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 = 49. # 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']
lr = 0.0625 * 0.01
weight_decay = 0
batch_size = 8
n_epochs = 100
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=batch_size,
shuffle=True)
# valid_loader = torch.utils.data.DataLoader(valid_set, batch_size=batch_size, shuffle=False)
test_loader = torch.utils.data.DataLoader(valid_set,
batch_size=batch_size,
shuffle=False)
# Send model to GPU
if cuda:
model.cuda(device)
from torch.optim import lr_scheduler
import torch.optim as optim
import argparse
parser = argparse.ArgumentParser(
description='cross subject domain adaptation')
parser.add_argument('--batch-size',
type=int,
default=50,
metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size',
type=int,
default=50,
metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs',
type=int,
default=100,
metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr',
type=float,
default=0.001,
metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum',
type=float,
default=0.5,
metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model',
action='store_true',
default=True,
help='For Saving the current Model')
args = parser.parse_args()
args.gpuidx = 0
args.seed = 0
args.use_tensorboard = False
args.save_model = False
optimizer = optim.AdamW(model.parameters(),
lr=lr,
weight_decay=weight_decay)
# scheduler = lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10, T_mult=1)
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=n_epochs - 1)
import pandas as pd
results_columns = ['test_loss', 'test_accuracy']
df = pd.DataFrame(columns=results_columns)
for epochidx in range(1, n_epochs):
print(epochidx)
train_crop(10, model, device, train_loader, optimizer, scheduler, cuda,
args.gpuidx)
test_loss, test_score = eval_crop(model, device, test_loader)
results = {'test_loss': test_loss, 'test_accuracy': test_score}
df = df.append(results, ignore_index=True)
print(results)
return df
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,
)
def run_exp_on_high_gamma_dataset(train_filename, test_filename, low_cut_hz,
model_name, max_epochs, max_increase_epochs,
np_th_seed, debug):
train_set, valid_set, test_set = load_train_valid_test(
train_filename=train_filename,
test_filename=test_filename,
low_cut_hz=low_cut_hz,
debug=debug)
if debug:
max_epochs = 4
set_random_seeds(np_th_seed, cuda=True)
#torch.backends.cudnn.benchmark = True# sometimes crashes?
n_classes = int(np.max(train_set.y) + 1)
n_chans = int(train_set.X.shape[1])
input_time_length = 1000
if model_name == 'deep':
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length=2).create_network()
elif model_name == 'shallow':
model = ShallowFBCSPNet(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length=30).create_network()
to_dense_prediction_model(model)
model.cuda()
model.eval()
out = model(np_to_var(train_set.X[:1, :, :input_time_length, None]).cuda())
n_preds_per_input = out.cpu().data.numpy().shape[2]
optimizer = optim.Adam(model.parameters(), weight_decay=0, lr=1e-3)
iterator = CropsFromTrialsIterator(batch_size=60,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input,
seed=np_th_seed)
monitors = [
LossMonitor(),
MisclassMonitor(col_suffix='sample_misclass'),
CroppedTrialMisclassMonitor(input_time_length=input_time_length),
RuntimeMonitor()
]
model_constraint = MaxNormDefaultConstraint()
loss_function = lambda preds, targets: F.nll_loss(th.mean(preds, dim=2),
targets)
run_after_early_stop = True
do_early_stop = True
remember_best_column = 'valid_misclass'
stop_criterion = Or([
MaxEpochs(max_epochs),
NoDecrease('valid_misclass', max_increase_epochs)
])
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator=iterator,
loss_function=loss_function,
optimizer=optimizer,
model_constraint=model_constraint,
monitors=monitors,
stop_criterion=stop_criterion,
remember_best_column=remember_best_column,
run_after_early_stop=run_after_early_stop,
cuda=True,
do_early_stop=do_early_stop)
exp.run()
return exp
def create_model(in_channels, num_classes, cuda=True):
def squeeze_out(x):
assert x.size()[1] == num_classes and x.size()[3] == 1
return x.squeeze(3).transpose(1, 2)
if cfg.TRAINING.MODEL.lower() == 'rnn':
model = RNNs(in_channels=in_channels)
elif 'deep4' in cfg.TRAINING.MODEL.lower():
if 'wide' in cfg.TRAINING.MODEL.lower():
pool_length = 4
pool_stride = 4
elif 'narrow' in cfg.TRAINING.MODEL.lower():
pool_length = 2
pool_stride = 2
else:
pool_length = 3
pool_stride = 3
model = Deep4Net(in_chans=in_channels,
n_classes=num_classes,
input_time_length=cfg.TRAINING.CROP_LEN,
pool_time_length=pool_length,
pool_time_stride=pool_stride,
final_conv_length=2,
stride_before_pool=True).create_network()
# remove softmax
new_model = nn.Sequential()
for name, module in model.named_children():
if name == 'softmax':
# continue
break
new_model.add_module(name, module)
# remove empty final dimension and permute output shape
new_model.add_module('squeeze', Expression(squeeze_out))
# if num_classes > 1:
# def transpose_class_time(x):
# return x.transpose(2, 1)
#
# new_model.add_module('trans', Expression(transpose_class_time))
model = new_model
to_dense_prediction_model(model)
elif cfg.TRAINING.MODEL.lower() == 'deep5':
# pool_time_length=3
# pool_time_stride=3
model = Deep5Net(in_chans=in_channels,
n_classes=num_classes,
input_time_length=cfg.TRAINING.CROP_LEN,
final_conv_length=2,
stride_before_pool=True).create_network()
# remove softmax
new_model = nn.Sequential()
for name, module in model.named_children():
if name == 'softmax':
# continue
break
new_model.add_module(name, module)
# remove empty final dimension and permute output shape
new_model.add_module('squeeze', Expression(squeeze_out))
# if num_classes > 1:
# def transpose_class_time(x):
# return x.transpose(2, 1)
#
# new_model.add_module('trans', Expression(transpose_class_time))
model = new_model
to_dense_prediction_model(model)
elif cfg.TRAINING.MODEL.lower() == 'shallow':
model = Shallow(in_chans=in_channels,
n_classes=num_classes,
input_time_length=cfg.TRAINING.CROP_LEN,
final_conv_length=2).create_network()
# remove softmax
new_model = nn.Sequential()
for name, module in model.named_children():
if name == 'softmax':
break
new_model.add_module(name, module)
# remove empty final dimension and permute output shape
new_model.add_module('squeeze', Expression(squeeze_out))
to_dense_prediction_model(model)
elif cfg.TRAINING.MODEL.lower() == 'hybrid':
model = Hybrid(in_channels=in_channels)
elif cfg.TRAINING.MODEL.lower() == 'tcn':
raise NotImplementedError
else:
assert False, f"Unknown Model {cfg.TRAINING.MODEL}"
optimizer = optim.Adam(model.parameters(),
lr=cfg.OPTIMIZATION.BASE_LR,
weight_decay=cfg.OPTIMIZATION.WEIGHT_DECAY)
scheduler = CosineAnnealingLR(optimizer, T_max=cfg.TRAINING.MAX_EPOCHS)
if cuda:
model.cuda()
model.eval()
metric = lambda targets, predictions: np.corrcoef(targets, predictions)[0,
1]
loss_fun = mse_loss
logger.info(model)
return model, optimizer, scheduler, loss_fun, metric
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,
)
def test_eeg_classifier():
# 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_window_samples = 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_window_samples=input_window_samples,
final_conv_length=12,
)
to_dense_prediction_model(model)
if cuda:
model.cuda()
# determine output size
test_input = np_to_var(
np.ones((2, in_chans, input_window_samples, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
train_set = create_from_X_y(X[:48],
y[:48],
drop_last_window=False,
window_size_samples=input_window_samples,
window_stride_samples=n_preds_per_input)
valid_set = create_from_X_y(X[48:60],
y[48:60],
drop_last_window=False,
window_size_samples=input_window_samples,
window_stride_samples=n_preds_per_input)
cropped_cb_train = CroppedTrialEpochScoring(
"accuracy",
name="train_trial_accuracy",
lower_is_better=False,
on_train=True,
)
cropped_cb_valid = CroppedTrialEpochScoring(
"accuracy",
on_train=False,
name="valid_trial_accuracy",
lower_is_better=False,
)
clf = EEGClassifier(
model,
criterion=CroppedLoss,
criterion__loss_function=nll_loss,
optimizer=optim.Adam,
train_split=predefined_split(valid_set),
batch_size=32,
callbacks=[
("train_trial_accuracy", cropped_cb_train),
("valid_trial_accuracy", cropped_cb_valid),
],
)
clf.fit(train_set, y=None, epochs=4)
expected = [
{
"batches": [
{
"train_loss": 1.9391239881515503,
"train_batch_size": 32
},
{
"train_loss": 2.895704507827759,
"train_batch_size": 32
},
{
"train_loss": 1.0713893175125122,
"train_batch_size": 32
},
{
"valid_loss": 1.1811838150024414,
"valid_batch_size": 24
},
],
"epoch":
1,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
1.9687392711639404,
"train_loss_best":
True,
"valid_loss":
1.1811838150024414,
"valid_loss_best":
True,
"train_trial_accuracy":
0.4791666666666667,
"train_trial_accuracy_best":
True,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
True,
},
{
"batches": [
{
"train_loss": 1.5488793849945068,
"train_batch_size": 32
},
{
"train_loss": 1.1174801588058472,
"train_batch_size": 32
},
{
"train_loss": 1.1525697708129883,
"train_batch_size": 32
},
{
"valid_loss": 2.202029228210449,
"valid_batch_size": 24
},
],
"epoch":
2,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
1.2729764382044475,
"train_loss_best":
True,
"valid_loss":
2.202029228210449,
"valid_loss_best":
False,
"train_trial_accuracy":
0.5,
"train_trial_accuracy_best":
True,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
False,
},
{
"batches": [
{
"train_loss": 1.0049529075622559,
"train_batch_size": 32
},
{
"train_loss": 1.0266971588134766,
"train_batch_size": 32
},
{
"train_loss": 1.0799436569213867,
"train_batch_size": 32
},
{
"valid_loss": 1.0638500452041626,
"valid_batch_size": 24
},
],
"epoch":
3,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
1.0371979077657063,
"train_loss_best":
True,
"valid_loss":
1.0638500452041626,
"valid_loss_best":
True,
"train_trial_accuracy":
0.5,
"train_trial_accuracy_best":
False,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
False,
},
{
"batches": [
{
"train_loss": 1.0052555799484253,
"train_batch_size": 32
},
{
"train_loss": 0.8479514718055725,
"train_batch_size": 32
},
{
"train_loss": 0.9589881300926208,
"train_batch_size": 32
},
{
"valid_loss": 0.8794112801551819,
"valid_batch_size": 24
},
],
"epoch":
4,
"train_batch_count":
3,
"valid_batch_count":
1,
"train_loss":
0.9373983939488729,
"train_loss_best":
True,
"valid_loss":
0.8794112801551819,
"valid_loss_best":
True,
"train_trial_accuracy":
0.5,
"train_trial_accuracy_best":
False,
"valid_trial_accuracy":
0.5,
"valid_trial_accuracy_best":
False,
},
]
history_without_dur = [{k: v
for k, v in h.items() if k != "dur"}
for h in clf.history]
assert_deep_allclose(expected, history_without_dur, atol=1e-3, rtol=1e-3)
def test_eeg_classifier():
# 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_window_samples = 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_window_samples=input_window_samples,
final_conv_length=12,
)
to_dense_prediction_model(model)
if cuda:
model.cuda()
# determine output size
test_input = np_to_var(
np.ones((2, in_chans, input_window_samples, 1), dtype=np.float32))
if cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
train_set = create_from_X_y(X[:48],
y[:48],
drop_last_window=False,
window_size_samples=input_window_samples,
window_stride_samples=n_preds_per_input)
valid_set = create_from_X_y(X[48:60],
y[48:60],
drop_last_window=False,
window_size_samples=input_window_samples,
window_stride_samples=n_preds_per_input)
cropped_cb_train = CroppedTrialEpochScoring(
"accuracy",
name="train_trial_accuracy",
lower_is_better=False,
on_train=True,
)
cropped_cb_valid = CroppedTrialEpochScoring(
"accuracy",
on_train=False,
name="valid_trial_accuracy",
lower_is_better=False,
)
clf = EEGClassifier(
model,
criterion=CroppedLoss,
criterion__loss_function=nll_loss,
optimizer=optim.Adam,
train_split=predefined_split(valid_set),
batch_size=32,
callbacks=[
("train_trial_accuracy", cropped_cb_train),
("valid_trial_accuracy", cropped_cb_valid),
],
)
clf.fit(train_set, y=None, epochs=4)
expected = [{
'batches': [{
'train_batch_size': 32,
'train_loss': 1.9391239881515503
}, {
'train_batch_size': 32,
'train_loss': 2.895704507827759
}, {
'train_batch_size': 32,
'train_loss': 1.0713887214660645
}, {
'valid_batch_size': 24,
'valid_loss': 1.18110191822052
}],
'epoch':
1,
'train_batch_count':
3,
'train_loss':
1.9687390724817913,
'train_loss_best':
True,
'train_trial_accuracy':
0.4791666666666667,
'train_trial_accuracy_best':
True,
'valid_batch_count':
1,
'valid_loss':
1.18110191822052,
'valid_loss_best':
True,
'valid_trial_accuracy':
0.5,
'valid_trial_accuracy_best':
True
}, {
'batches': [{
'train_batch_size': 32,
'train_loss': 1.6741573810577393
}, {
'train_batch_size': 32,
'train_loss': 0.9984264373779297
}, {
'train_batch_size': 32,
'train_loss': 1.1340471506118774
}, {
'valid_batch_size': 24,
'valid_loss': 2.5375664234161377
}],
'epoch':
2,
'train_batch_count':
3,
'train_loss':
1.2688769896825154,
'train_loss_best':
True,
'train_trial_accuracy':
0.5,
'train_trial_accuracy_best':
True,
'valid_batch_count':
1,
'valid_loss':
2.5375664234161377,
'valid_loss_best':
False,
'valid_trial_accuracy':
0.5,
'valid_trial_accuracy_best':
False
}, {
'batches': [{
'train_batch_size': 32,
'train_loss': 0.8795645833015442
}, {
'train_batch_size': 32,
'train_loss': 1.0339491367340088
}, {
'train_batch_size': 32,
'train_loss': 1.19275963306427
}, {
'valid_batch_size': 24,
'valid_loss': 1.655737042427063
}],
'epoch':
3,
'train_batch_count':
3,
'train_loss':
1.0354244510332744,
'train_loss_best':
True,
'train_trial_accuracy':
0.5,
'train_trial_accuracy_best':
False,
'valid_batch_count':
1,
'valid_loss':
1.655737042427063,
'valid_loss_best':
False,
'valid_trial_accuracy':
0.5,
'valid_trial_accuracy_best':
False
}, {
'batches': [{
'train_batch_size': 32,
'train_loss': 1.1963350772857666
}, {
'train_batch_size': 32,
'train_loss': 0.8621770143508911
}, {
'train_batch_size': 32,
'train_loss': 1.099318265914917
}, {
'valid_batch_size': 24,
'valid_loss': 1.0293445587158203
}],
'epoch':
4,
'train_batch_count':
3,
'train_loss':
1.0526101191838582,
'train_loss_best':
False,
'train_trial_accuracy':
0.625,
'train_trial_accuracy_best':
True,
'valid_batch_count':
1,
'valid_loss':
1.0293445587158203,
'valid_loss_best':
True,
'valid_trial_accuracy':
0.25,
'valid_trial_accuracy_best':
False
}]
history_without_dur = [{k: v
for k, v in h.items() if k != "dur"}
for h in clf.history]
assert_deep_allclose(expected, history_without_dur, atol=1e-3, rtol=1e-3)
def exp(subject_id):
import torch
input_window_samples = 1000
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.deterministic = True
torch.backends.cudnn.benchmark = False
seed = 20190706 # random seed to make results reproducible
# Set random seed to be able to reproduce results
random.seed(seed)
torch.manual_seed(seed)
if cuda:
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
n_classes = 4
PATH = '../datasets/'
with open(PATH + 'bcic_datasets_[0,49].pkl', 'rb') as f:
data = pickle.load(f)
import torch
print('subject:' + str(subject_id))
#make train test
tr = []
val =[]
test_train_split = 0.5
dataset= data[subject_id]
dataset_size = len(dataset)
indices = list(range(dataset_size))
test_split = int(np.floor(test_train_split * dataset_size))
train_indices, test_indices = indices[:test_split], indices[test_split:]
np.random.shuffle(train_indices)
#분석
sample_data = data[0].dataset
sample_data.psd()
from mne.viz import plot_epochs_image
import mne
plot_epochs_image(sample_data, picks=['C3','C4'])
label = sample_data.read_label()
sample_data.plot_projs_topomap()
train_sampler = SubsetRandomSampler(train_indices)
test_sampler = SubsetRandomSampler(test_indices)
from braindecode.models import ShallowFBCSPNet
model = ShallowFBCSPNet(
22,
n_classes,
input_window_samples=input_window_samples,
final_conv_length=30,
)
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, 22, input_window_samples)[2]
print("n_preds_per_input : ", n_preds_per_input)
print(model)
# crop_size =1000
#
#
#
#
# model = ShallowNet_dense(n_classes, 22, crop_size)
#
# print(model)
epochs = 100
# For deep4 they should be:
lr = 1 * 0.01
weight_decay = 0.5 * 0.001
batch_size = 8
train_set = torch.utils.data.Subset(dataset,indices= train_indices)
test_set = torch.utils.data.Subset(dataset,indices= test_indices)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=False)
# Send model to GPU
if cuda:
model.cuda(device=device)
from torch.optim import lr_scheduler
import torch.optim as optim
import argparse
parser = argparse.ArgumentParser(description='cross subject domain adaptation')
parser.add_argument('--batch-size', type=int, default=50, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=50, metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=100, metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr', type=float, default=0.001, metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model', action='store_true', default=True,
help='For Saving the current Model')
args = parser.parse_args()
args.gpuidx = 0
args.seed = 0
args.use_tensorboard = False
args.save_model = False
lr = 0.0625 * 0.01
weight_decay = 0
optimizer = optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
# scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=200)
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max = epochs-1)
#
# #test lr
# lr = []
# for i in range(200):
# scheduler.step()
# lr.append(scheduler.get_lr())
#
# import matplotlib.pyplot as plt
# plt.plot(lr)
import pandas as pd
results_columns = ['test_loss', 'test_accuracy']
df = pd.DataFrame(columns=results_columns)
for epochidx in range(1, epochs):
print(epochidx)
train_crop(10, model, device, train_loader,optimizer,scheduler,cuda, args.gpuidx)
test_loss, test_score = eval_crop(model, device, test_loader)
results = {'test_loss': test_loss, 'test_accuracy': test_score}
df = df.append(results, ignore_index=True)
print(results)
return df
optimizer_lr = 0.01
optimizer_weight_decay = 0.0005
else:
raise ValueError(f'{model_name} unknown')
new_model = torch.nn.Sequential()
for name, module_ in model.named_children():
if "softmax" in name:
continue
new_model.add_module(name, module_)
model = new_model
if cuda:
model.cuda()
to_dense_prediction_model(model)
n_preds_per_input = get_output_shape(model, n_chans, input_window_samples)[2]
train_set, valid_set = create_compatible_dataset('./data/BCICIV_4_mat/sub1_comp.mat')
# dataset = create_fixed_length_windows(
# dataset,
# start_offset_samples=0,
# stop_offset_samples=0,
# window_size_samples=input_window_samples,
# window_stride_samples=n_preds_per_input,
# drop_last_window=False,
# drop_bad_windows=True,
# )
# splits = dataset.split("session")
def run_exp(data_folder, subject_id, low_cut_hz, model, cuda):
train_filename = 'A{:01d}T.gdf'.format(subject_id)
test_filename = 'A{:01d}E.gdf'.format(subject_id)
train_filepath = os.path.join(data_folder, train_filename)
test_filepath = os.path.join(data_folder, test_filename)
train_label_filepath = train_filepath.replace('.gdf', '.mat')
test_label_filepath = test_filepath.replace('.gdf', '.mat')
train_loader = BCICompetition4Set2A(
train_filepath, labels_filename=train_label_filepath)
test_loader = BCICompetition4Set2A(
test_filepath, labels_filename=test_label_filepath)
#print(train_loader)
train_cnt = train_loader.load()
test_cnt = test_loader.load()
# Preprocessing
train_cnt = train_cnt.drop_channels(['STI 014', 'EOG-left',
'EOG-central', 'EOG-right'])
assert len(train_cnt.ch_names) == 22
# lets convert to millvolt for numerical stability of next operations
train_cnt = mne_apply(lambda a: a * 1e6, train_cnt)
train_cnt = mne_apply(
lambda a: bandpass_cnt(a, low_cut_hz, 38, train_cnt.info['sfreq'],
filt_order=3,
axis=1), train_cnt)
train_cnt = mne_apply(
lambda a: exponential_running_standardize(a.T, factor_new=1e-3,
init_block_size=1000,
eps=1e-4).T,
train_cnt)
test_cnt = test_cnt.drop_channels(['STI 014', 'EOG-left',
'EOG-central', 'EOG-right'])
assert len(test_cnt.ch_names) == 22
test_cnt = mne_apply(lambda a: a * 1e6, test_cnt)
test_cnt = mne_apply(
lambda a: bandpass_cnt(a, low_cut_hz, 38, test_cnt.info['sfreq'],
filt_order=3,
axis=1), test_cnt)
test_cnt = mne_apply(
lambda a: exponential_running_standardize(a.T, factor_new=1e-3,
init_block_size=1000,
eps=1e-4).T,
test_cnt)
marker_def = OrderedDict([('Left Hand', [1]), ('Right Hand', [2],),
('Foot', [3]), ('Tongue', [4])])
ival = [-500, 4000]
train_set = create_signal_target_from_raw_mne(train_cnt, marker_def, ival)
test_set = create_signal_target_from_raw_mne(test_cnt, marker_def, ival)
train_set, valid_set = split_into_two_sets(train_set,
first_set_fraction=0.8)
set_random_seeds(seed=20190706, cuda=cuda)
n_classes = 4
n_chans = int(train_set.X.shape[1])
input_time_length=1000
if model == 'shallow':
model = ShallowFBCSPNet(n_chans, n_classes, input_time_length=input_time_length,
final_conv_length=30).create_network()
elif model == 'deep':
model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length,
final_conv_length=2).create_network()
to_dense_prediction_model(model)
if cuda:
model.cuda()
log.info("Model: \n{:s}".format(str(model)))
dummy_input = np_to_var(train_set.X[:1, :, :, None])
if cuda:
dummy_input = dummy_input.cuda()
out = model(dummy_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
optimizer = optim.Adam(model.parameters())
iterator = CropsFromTrialsIterator(batch_size=60,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
stop_criterion = Or([MaxEpochs(800),
NoDecrease('valid_misclass', 80)])
monitors = [LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'),
CroppedTrialMisclassMonitor(
input_time_length=input_time_length), RuntimeMonitor()]
model_constraint = MaxNormDefaultConstraint()
loss_function = lambda preds, targets: F.nll_loss(
th.mean(preds, dim=2, keepdim=False), targets)
exp = Experiment(model, train_set, valid_set, test_set, iterator=iterator,
loss_function=loss_function, optimizer=optimizer,
model_constraint=model_constraint,
monitors=monitors,
stop_criterion=stop_criterion,
remember_best_column='valid_misclass',
run_after_early_stop=True, cuda=cuda)
exp.run()
return exp