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 train_model_once(subject_id, i_valid_fold, config, model_state_dict=None):
# Data loading
data = get_dataset(subject_id, i_valid_fold,
config["experiment"]["dataset"], config)
# import pickle
# from base.base_data_loader import BaseDataLoader
# from braindecode.datautil.splitters import split_into_train_valid_test
# pickle_path = os.path.abspath(
# os.path.join(os.path.dirname(__file__), '..',
# 'data/bcic_iv_2a_all_9_subjects.pickle'))
# with open(pickle_path, 'rb') as f:
# data = pickle.load(f)
# data = data[0]
# train, valid, test = split_into_train_valid_test(data, 4, 0)
# data = BaseDataLoader(train, valid, test, 4)
# Build model architecture
model = get_model(data, model_state_dict, config)
# Set iterator and metric function handle
iterator = get_iterator(model, data, config)
predict_label_func = get_prediction_func(config)
# Get function handle of loss
loss_function = get_loss(config)
# Build optimizer, learning rate scheduler
stop_criterion = get_stop_criterion(config)
optimizer = get_optmizer(model, config)
print(model)
# Init trainer and train
trainer = Trainer(
data.train_set,
data.validation_set,
data.test_set,
model,
optimizer,
iterator,
loss_function,
stop_criterion,
model_constraint=MaxNormDefaultConstraint(),
cuda=torch.cuda.is_available(),
func_compute_pred_labels=predict_label_func,
siamese=(config["experiment"]["type"] == "ccsa_da"),
)
trainer.train()
# Save results
log_training_results(trainer)
file_state_dict = save_result_and_model(trainer, model, config)
return file_state_dict
def train(train_set, test_set, model, iterator, monitors, loss_function,
max_epochs, cuda):
if cuda:
model.cuda()
optimizer = AdamW(model.parameters(),
lr=1 * 0.01,
weight_decay=0.5 *
0.001) # these are good values for the deep model
stop_criterion = MaxEpochs(max_epochs)
model_constraint = MaxNormDefaultConstraint()
n_updates_per_epoch = sum(
[1 for _ in iterator.get_batches(train_set, shuffle=True)])
n_updates_per_period = n_updates_per_epoch * max_epochs
scheduler = CosineAnnealing(n_updates_per_period)
optimizer = ScheduledOptimizer(scheduler,
optimizer,
schedule_weight_decay=True)
exp = Experiment(model,
train_set,
None,
test_set,
iterator=iterator,
loss_function=loss_function,
optimizer=optimizer,
model_constraint=model_constraint,
monitors=monitors,
remember_best_column=None,
stop_criterion=stop_criterion,
cuda=cuda,
run_after_early_stop=False,
do_early_stop=False)
exp.run()
return exp
def build_exp(model_name, cuda, data, batch_size, max_epochs, max_increase_epochs):
log.info("==============================")
log.info("Loading Data...")
log.info("==============================")
train_set = data.train_set
valid_set = data.validation_set
test_set = data.test_set
log.info("==============================")
log.info("Setting Up Model...")
log.info("==============================")
set_random_seeds(seed=20190706, cuda=cuda)
n_classes = 4
n_chans = int(train_set.X.shape[1])
input_time_length = train_set.X.shape[2]
if model_name == "shallow":
model = NewShallowNet(
n_chans, n_classes, input_time_length, final_conv_length="auto"
)
# model = ShallowFBCSPNet(
# n_chans,
# n_classes,
# input_time_length=input_time_length,
# final_conv_length="auto",
# ).create_network()
elif model_name == "deep":
model = NewDeep4Net(n_chans, n_classes, input_time_length, "auto")
# model = Deep4Net(
# n_chans,
# n_classes,
# input_time_length=input_time_length,
# final_conv_length="auto",
# ).create_network()
elif model_name == "eegnet":
# model = EEGNet(n_chans, n_classes,
# input_time_length=input_time_length)
# model = EEGNetv4(n_chans, n_classes,
# input_time_length=input_time_length).create_network()
model = NewEEGNet(n_chans, n_classes, input_time_length=input_time_length)
if cuda:
model.cuda()
log.info("==============================")
log.info("Logging Model Architecture:")
log.info("==============================")
log.info("Model: \n{:s}".format(str(model)))
log.info("==============================")
log.info("Building Experiment:")
log.info("==============================")
optimizer = optim.Adam(model.parameters())
iterator = BalancedBatchSizeIterator(batch_size=batch_size)
stop_criterion = Or(
[MaxEpochs(max_epochs), NoDecrease("valid_misclass", max_increase_epochs)]
)
monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()]
model_constraint = MaxNormDefaultConstraint()
exp = Experiment(
model,
train_set,
valid_set,
test_set,
iterator=iterator,
loss_function=F.nll_loss,
optimizer=optimizer,
model_constraint=model_constraint,
monitors=monitors,
stop_criterion=stop_criterion,
remember_best_column="valid_misclass",
run_after_early_stop=True,
cuda=cuda,
)
return exp
def run_exp(epoches, batch_size, subject_num, model_type, cuda, single_subject,
single_subject_num):
# ival = [-500, 4000]
max_increase_epochs = 160
# Preprocessing
X, y = loadSubjects(subject_num, single_subject, single_subject_num)
X = X.astype(np.float32)
y = y.astype(np.int64)
X, y = shuffle(X, y)
trial_length = X.shape[2]
print("trial_length " + str(trial_length))
print("trying to run with {} sec trials ".format((trial_length - 1) / 256))
print("y")
print(y)
trainingSampleSize = int(len(X) * 0.6)
valudationSampleSize = int(len(X) * 0.2)
testSampleSize = int(len(X) * 0.2)
print("INFO : Training sample size: {}".format(trainingSampleSize))
print("INFO : Validation sample size: {}".format(valudationSampleSize))
print("INFO : Test sample size: {}".format(testSampleSize))
train_set = SignalAndTarget(X[:trainingSampleSize],
y=y[:trainingSampleSize])
valid_set = SignalAndTarget(
X[trainingSampleSize:(trainingSampleSize + valudationSampleSize)],
y=y[trainingSampleSize:(trainingSampleSize + valudationSampleSize)])
test_set = SignalAndTarget(X[(trainingSampleSize + valudationSampleSize):],
y=y[(trainingSampleSize +
valudationSampleSize):])
set_random_seeds(seed=20190706, cuda=cuda)
n_classes = 3
n_chans = int(train_set.X.shape[1])
input_time_length = train_set.X.shape[2]
if model_type == 'shallow':
model = ShallowFBCSPNet(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto').create_network()
elif model_type == 'deep':
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto').create_network()
elif model_type == 'eegnet':
model = EEGNetv4(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto').create_network()
if cuda:
model.cuda()
log.info("Model: \n{:s}".format(str(model)))
optimizer = optim.Adam(model.parameters())
iterator = BalancedBatchSizeIterator(batch_size=batch_size)
stop_criterion = Or([
MaxEpochs(max_epochs),
NoDecrease('valid_misclass', max_increase_epochs)
])
monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()]
model_constraint = MaxNormDefaultConstraint()
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator=iterator,
loss_function=F.nll_loss,
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()
# th.save(model, "models\{}-cropped-singleSubjectNum{}-{}sec-{}epoches-torch_model".format(model_type, single_subject_num, ((trial_length - 1) / 256), epoches))
return exp
level=logging.INFO, stream=sys.stdout)
WINDOW_LEN = 200
OVERLAP = 150
windows = windows_index(500,WINDOW_LEN,OVERLAP,250)
hyp_params = dict(window=windows[:2],
activation=["leaky_relu"],
structure= ["shallow"])
parameters = dict(best_loss = 100.0,
batch_size = 32,
monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()],
model_constraint = MaxNormDefaultConstraint(),
max_increase_epochs = 0,
cuda = True,
epochs=1,
learning_rate_scheduler=StepLR,
lr_step=20, lr_gamma=0.9)
EEGSubNet_params = dict(n_filters_time=40, filter_time_length=5, n_filters_spat=40, n_filters_2=20, filter_length_2=20,
pool_time_length_1=5, pool_time_stride_1=2, pool_length_2=5, pool_stride_2=3, final_conv_length='auto',
conv_nonlin=th.nn.functional.leaky_relu, pool_mode='mean', pool_nonlin=safe_log,
split_first_layer=True, batch_norm=True, batch_norm_alpha=0.2,
drop_prob=0.1)
fNIRSSubNet_params = dict(n_filters_time=40, filter_time_length=5, n_filters_spat=40, n_filters_2=20, filter_length_2=20,
def run_exp(data_folder, subject_id, low_cut_hz, model, cuda):
train_filename = 'A{:02d}T.gdf'.format(subject_id)
test_filename = 'A{:02d}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)
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 = train_set.X.shape[2]
if model == 'shallow':
model = ShallowFBCSPNet(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto').create_network()
elif model == 'deep':
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto').create_network()
if cuda:
model.cuda()
log.info("Model: \n{:s}".format(str(model)))
optimizer = optim.Adam(model.parameters())
iterator = BalancedBatchSizeIterator(batch_size=60)
stop_criterion = Or([MaxEpochs(1600), NoDecrease('valid_misclass', 160)])
monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()]
model_constraint = MaxNormDefaultConstraint()
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator=iterator,
loss_function=F.nll_loss,
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
def run_exp(max_epochs, only_return_exp):
from collections import OrderedDict
filenames = [
'data/robot-hall/NiRiNBD6.ds_1-1_500Hz.BBCI.mat',
'data/robot-hall/NiRiNBD8.ds_1-1_500Hz.BBCI.mat',
'data/robot-hall/NiRiNBD9.ds_1-1_500Hz.BBCI.mat',
'data/robot-hall/NiRiNBD10.ds_1-1_500Hz.BBCI.mat',
'data/robot-hall/NiRiNBD12_cursor_250Hz.BBCI.mat',
'data/robot-hall/NiRiNBD13_cursorS000R01_onlyFullRuns_250Hz.BBCI.mat',
'data/robot-hall/NiRiNBD14_cursor_250Hz.BBCI.mat',
'data/robot-hall/NiRiNBD15_cursor_250Hz.BBCI.mat'
]
sensor_names = [
'Fp1', 'Fpz', 'Fp2', 'AF7', 'AF3', 'AF4', 'AF8', 'F7', 'F5', 'F3',
'F1', 'Fz', 'F2', 'F4', 'F6', 'F8', 'FT7', 'FC5', 'FC3', 'FC1', 'FCz',
'FC2', 'FC4', 'FC6', 'FT8', 'M1', 'T7', 'C5', 'C3', 'C1', 'Cz', 'C2',
'C4', 'C6', 'T8', 'M2', 'TP7', 'CP5', 'CP3', 'CP1', 'CPz', 'CP2',
'CP4', 'CP6', 'TP8', 'P7', 'P5', 'P3', 'P1', 'Pz', 'P2', 'P4', 'P6',
'P8', 'PO7', 'PO5', 'PO3', 'POz', 'PO4', 'PO6', 'PO8', 'O1', 'Oz', 'O2'
]
name_to_start_codes = OrderedDict([('Right Hand', [1]), ('Feet', [2]),
('Rotation', [3]), ('Words', [4])])
name_to_stop_codes = OrderedDict([('Right Hand', [10]), ('Feet', [20]),
('Rotation', [30]), ('Words', [40])])
trial_ival = [500, 0]
min_break_length_ms = 6000
max_break_length_ms = 8000
break_ival = [1000, -500]
input_time_length = 700
filename_to_extra_args = {
'data/robot-hall/NiRiNBD12_cursor_250Hz.BBCI.mat':
dict(
name_to_start_codes=OrderedDict([('Right Hand', [1]),
('Feet', [2]), ('Rotation', [3]),
('Words', [4]), ('Rest', [5])]),
name_to_stop_codes=OrderedDict([('Right Hand', [10]),
('Feet', [20]), ('Rotation', [30]),
('Words', [40]), ('Rest', [50])]),
min_break_length_ms=3700,
max_break_length_ms=3900,
),
'data/robot-hall/NiRiNBD13_cursorS000R01_onlyFullRuns_250Hz.BBCI.mat':
dict(
name_to_start_codes=OrderedDict([('Right Hand', [1]),
('Feet', [2]), ('Rotation', [3]),
('Words', [4]), ('Rest', [5])]),
name_to_stop_codes=OrderedDict([('Right Hand', [10]),
('Feet', [20]), ('Rotation', [30]),
('Words', [40]), ('Rest', [50])]),
min_break_length_ms=3700,
max_break_length_ms=3900,
),
'data/robot-hall/NiRiNBD14_cursor_250Hz.BBCI.mat':
dict(
name_to_start_codes=OrderedDict([('Right Hand', [1]),
('Feet', [2]), ('Rotation', [3]),
('Words', [4]), ('Rest', [5])]),
name_to_stop_codes=OrderedDict([('Right Hand', [10]),
('Feet', [20]), ('Rotation', [30]),
('Words', [40]), ('Rest', [50])]),
min_break_length_ms=3700,
max_break_length_ms=3900,
),
'data/robot-hall/NiRiNBD15_cursor_250Hz.BBCI.mat':
dict(
name_to_start_codes=OrderedDict([('Right Hand', [1]),
('Feet', [2]), ('Rotation', [3]),
('Words', [4]), ('Rest', [5])]),
name_to_stop_codes=OrderedDict([('Right Hand', [10]),
('Feet', [20]), ('Rotation', [30]),
('Words', [40]), ('Rest', [50])]),
min_break_length_ms=3700,
max_break_length_ms=3900,
),
}
from braindecode.datautil.trial_segment import \
create_signal_target_with_breaks_from_mne
from copy import deepcopy
def load_data(filenames, sensor_names, name_to_start_codes,
name_to_stop_codes, trial_ival, break_ival,
min_break_length_ms, max_break_length_ms, input_time_length,
filename_to_extra_args):
all_sets = []
original_args = locals()
for filename in filenames:
kwargs = deepcopy(original_args)
if filename in filename_to_extra_args:
kwargs.update(filename_to_extra_args[filename])
log.info("Loading {:s}...".format(filename))
cnt = BBCIDataset(filename, load_sensor_names=sensor_names).load()
cnt = cnt.drop_channels(['STI 014'])
log.info("Resampling...")
cnt = resample_cnt(cnt, 100)
log.info("Standardizing...")
cnt = mne_apply(
lambda a: exponential_running_standardize(
a.T, init_block_size=50).T, cnt)
log.info("Transform to set...")
full_set = (create_signal_target_with_breaks_from_mne(
cnt,
kwargs['name_to_start_codes'],
kwargs['trial_ival'],
kwargs['name_to_stop_codes'],
kwargs['min_break_length_ms'],
kwargs['max_break_length_ms'],
kwargs['break_ival'],
prepad_trials_to_n_samples=kwargs['input_time_length'],
))
all_sets.append(full_set)
return all_sets
sensor_names = [
'Fp1', 'Fpz', 'Fp2', 'AF7', 'AF3', 'AF4', 'AF8', 'F7', 'F5', 'F3',
'F1', 'Fz', 'F2', 'F4', 'F6', 'F8', 'FT7', 'FC5', 'FC3', 'FC1', 'FCz',
'FC2', 'FC4', 'FC6', 'FT8', 'M1', 'T7', 'C5', 'C3', 'C1', 'Cz', 'C2',
'C4', 'C6', 'T8', 'M2', 'TP7', 'CP5', 'CP3', 'CP1', 'CPz', 'CP2',
'CP4', 'CP6', 'TP8', 'P7', 'P5', 'P3', 'P1', 'Pz', 'P2', 'P4', 'P6',
'P8', 'PO7', 'PO5', 'PO3', 'POz', 'PO4', 'PO6', 'PO8', 'O1', 'Oz', 'O2'
]
#sensor_names = ['C3', 'C4']
n_chans = len(sensor_names)
if not only_return_exp:
all_sets = load_data(filenames, sensor_names, name_to_start_codes,
name_to_stop_codes, trial_ival, break_ival,
min_break_length_ms, max_break_length_ms,
input_time_length, filename_to_extra_args)
from braindecode.datautil.signal_target import SignalAndTarget
from braindecode.datautil.splitters import concatenate_sets
train_set = concatenate_sets(all_sets[:6])
valid_set = all_sets[6]
test_set = all_sets[7]
else:
train_set = None
valid_set = None
test_set = None
set_random_seeds(seed=20171017, cuda=True)
n_classes = 5
# final_conv_length determines the size of the receptive field of the ConvNet
model = ShallowFBCSPNet(in_chans=n_chans,
n_classes=n_classes,
input_time_length=input_time_length,
final_conv_length=30).create_network()
to_dense_prediction_model(model)
model.cuda()
from torch import optim
import numpy as np
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, n_chans, input_time_length, 1), dtype=np.float32))
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.experiments.experiment import Experiment
from braindecode.experiments.monitors import RuntimeMonitor, LossMonitor, \
CroppedTrialMisclassMonitor, MisclassMonitor
from braindecode.experiments.stopcriteria import MaxEpochs
from braindecode.torch_ext.losses import log_categorical_crossentropy
import torch.nn.functional as F
import torch as th
from braindecode.torch_ext.modules import Expression
loss_function = log_categorical_crossentropy
model_constraint = MaxNormDefaultConstraint()
monitors = [
LossMonitor(),
MisclassMonitor(col_suffix='sample_misclass'),
CroppedTrialMisclassMonitor(input_time_length),
RuntimeMonitor(),
]
stop_criterion = MaxEpochs(max_epochs)
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator,
loss_function,
optimizer,
model_constraint,
monitors,
stop_criterion,
remember_best_column='valid_sample_misclass',
run_after_early_stop=True,
batch_modifier=None,
cuda=True)
if not only_return_exp:
exp.run()
return exp
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_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 run_exp(data_folder, subject_id, low_cut_hz, model, cuda):
ival = [-500, 4000]
max_epochs = 1600
max_increase_epochs = 160
batch_size = 60
high_cut_hz = 38
factor_new = 1e-3
init_block_size = 1000
valid_set_fraction = 0.2
train_filename = "A{:02d}T.gdf".format(subject_id)
test_filename = "A{:02d}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
)
train_cnt = train_loader.load()
test_cnt = test_loader.load()
# Preprocessing
train_cnt = train_cnt.drop_channels(
["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,
high_cut_hz,
train_cnt.info["sfreq"],
filt_order=3,
axis=1,
),
train_cnt,
)
train_cnt = mne_apply(
lambda a: exponential_running_standardize(
a.T,
factor_new=factor_new,
init_block_size=init_block_size,
eps=1e-4,
).T,
train_cnt,
)
test_cnt = test_cnt.drop_channels(["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,
high_cut_hz,
test_cnt.info["sfreq"],
filt_order=3,
axis=1,
),
test_cnt,
)
test_cnt = mne_apply(
lambda a: exponential_running_standardize(
a.T,
factor_new=factor_new,
init_block_size=init_block_size,
eps=1e-4,
).T,
test_cnt,
)
marker_def = OrderedDict(
[
("Left Hand", [1]),
("Right Hand", [2]),
("Foot", [3]),
("Tongue", [4]),
]
)
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=1 - valid_set_fraction
)
set_random_seeds(seed=20190706, cuda=cuda)
n_classes = 4
n_chans = int(train_set.X.shape[1])
input_time_length = train_set.X.shape[2]
if model == "shallow":
model = ShallowFBCSPNet(
n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length="auto",
).create_network()
elif model == "deep":
model = Deep4Net(
n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length="auto",
).create_network()
if cuda:
model.cuda()
log.info("Model: \n{:s}".format(str(model)))
optimizer = optim.Adam(model.parameters())
iterator = BalancedBatchSizeIterator(batch_size=batch_size)
stop_criterion = Or(
[
MaxEpochs(max_epochs),
NoDecrease("valid_misclass", max_increase_epochs),
]
)
monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()]
model_constraint = MaxNormDefaultConstraint()
exp = Experiment(
model,
train_set,
valid_set,
test_set,
iterator=iterator,
loss_function=F.nll_loss,
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
def network_model(subject_id, model_type, data_type, cropped, cuda, parameters, hyp_params):
best_params = dict() # dictionary to store hyper-parameter values
#####Parameter passed to funciton#####
max_epochs = parameters['max_epochs']
max_increase_epochs = parameters['max_increase_epochs']
batch_size = parameters['batch_size']
#####Constant Parameters#####
best_loss = 100.0 # instatiate starting point for loss
iterator = BalancedBatchSizeIterator(batch_size=batch_size)
stop_criterion = Or([MaxEpochs(max_epochs),
NoDecrease('valid_misclass', max_increase_epochs)])
monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()]
model_constraint = MaxNormDefaultConstraint()
epoch = 4096
#####Collect and format data#####
if data_type == 'words':
data, labels = format_data(data_type, subject_id, epoch)
data = data[:,:,768:1280] # within-trial window selected for classification
elif data_type == 'vowels':
data, labels = format_data(data_type, subject_id, epoch)
data = data[:,:,512:1024]
elif data_type == 'all_classes':
data, labels = format_data(data_type, subject_id, epoch)
data = data[:,:,768:1280]
x = lambda a: a * 1e6 # improves numerical stability
data = x(data)
data = normalize(data)
data, labels = balanced_subsample(data, labels) # downsampling the data to ensure equal classes
data, _, labels, _ = train_test_split(data, labels, test_size=0, random_state=42) # redundant shuffle of data/labels
#####model inputs#####
unique, counts = np.unique(labels, return_counts=True)
n_classes = len(unique)
n_chans = int(data.shape[1])
input_time_length = data.shape[2]
#####k-fold nested corss-validation#####
num_folds = 4
skf = StratifiedKFold(n_splits=num_folds, shuffle=True, random_state=10)
out_fold_num = 0 # outer-fold number
cv_scores = []
#####Outer=Fold#####
for inner_ind, outer_index in skf.split(data, labels):
inner_fold, outer_fold = data[inner_ind], data[outer_index]
inner_labels, outer_labels = labels[inner_ind], labels[outer_index]
out_fold_num += 1
# list for storing cross-validated scores
loss_with_params = dict()# for storing param values and losses
in_fold_num = 0 # inner-fold number
#####Inner-Fold#####
for train_idx, valid_idx in skf.split(inner_fold, inner_labels):
X_Train, X_val = inner_fold[train_idx], inner_fold[valid_idx]
y_train, y_val = inner_labels[train_idx], inner_labels[valid_idx]
in_fold_num += 1
train_set = SignalAndTarget(X_Train, y_train)
valid_set = SignalAndTarget(X_val, y_val)
loss_with_params[f"Fold_{in_fold_num}"] = dict()
####Nested cross-validation#####
for drop_prob in hyp_params['drop_prob']:
for loss_function in hyp_params['loss']:
for i in range(len(hyp_params['lr_adam'])):
model = None # ensure no duplication of models
# model, learning-rate and optimizer setup according to model_type
if model_type == 'shallow':
model = ShallowFBCSPNet(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length,
n_filters_time=80, filter_time_length=40, n_filters_spat=80,
pool_time_length=75, pool_time_stride=25, final_conv_length='auto',
conv_nonlin=square, pool_mode='max', pool_nonlin=safe_log,
split_first_layer=True, batch_norm=True, batch_norm_alpha=0.1,
drop_prob=drop_prob).create_network()
lr = hyp_params['lr_ada'][i]
optimizer = optim.Adadelta(model.parameters(), lr=lr, rho=0.9, weight_decay=0.1, eps=1e-8)
elif model_type == 'deep':
model = Deep4Net(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length,
final_conv_length='auto', n_filters_time=20, n_filters_spat=20, filter_time_length=10,
pool_time_length=3, pool_time_stride=3, n_filters_2=50, filter_length_2=15,
n_filters_3=100, filter_length_3=15, n_filters_4=400, filter_length_4=10,
first_nonlin=leaky_relu, first_pool_mode='max', first_pool_nonlin=safe_log, later_nonlin=leaky_relu,
later_pool_mode='max', later_pool_nonlin=safe_log, drop_prob=drop_prob,
double_time_convs=False, split_first_layer=False, batch_norm=True, batch_norm_alpha=0.1,
stride_before_pool=False).create_network() #filter_length_4 changed from 15 to 10
lr = hyp_params['lr_ada'][i]
optimizer = optim.Adadelta(model.parameters(), lr=lr, weight_decay=0.1, eps=1e-8)
elif model_type == 'eegnet':
model = EEGNetv4(in_chans=n_chans, n_classes=n_classes, final_conv_length='auto',
input_time_length=input_time_length, pool_mode='mean', F1=16, D=2, F2=32,
kernel_length=64, third_kernel_size=(8,4), drop_prob=drop_prob).create_network()
lr = hyp_params['lr_adam'][i]
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=0, eps=1e-8, amsgrad=False)
set_random_seeds(seed=20190629, cuda=cuda)
if cuda:
model.cuda()
torch.backends.cudnn.deterministic = True
model = torch.nn.DataParallel(model)
log.info("%s model: ".format(str(model)))
loss_function = loss_function
model_loss_function = None
#####Setup to run the selected model#####
model_test = Experiment(model, train_set, valid_set, test_set=None, 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, model_loss_function=model_loss_function, cuda=cuda,
data_type=data_type, subject_id=subject_id, model_type=model_type,
cropped=cropped, model_number=str(out_fold_num))
model_test.run()
model_loss = model_test.epochs_df['valid_loss'].astype('float')
current_val_loss = current_loss(model_loss)
loss_with_params[f"Fold_{in_fold_num}"][f"{drop_prob}/{loss_function}/{lr}"] = current_val_loss
####Select and train optimized model#####
df = pd.DataFrame(loss_with_params)
df['mean'] = df.mean(axis=1) # compute mean loss across k-folds
writer_df = f"results_folder\\results\\S{subject_id}\\{model_type}_parameters.xlsx"
df.to_excel(writer_df)
best_dp, best_loss, best_lr = df.loc[df['mean'].idxmin()].__dict__['_name'].split("/") # extract best param values
if str(best_loss[10:13]) == 'nll':
best_loss = F.nll_loss
elif str(best_loss[10:13]) == 'cro':
best_loss = F.cross_entropy
print(f"Best parameters: dropout: {best_dp}, loss: {str(best_loss)[10:13]}, lr: {best_lr}")
#####Train model on entire inner fold set#####
torch.backends.cudnn.deterministic = True
model = None
#####Create outer-fold validation and test sets#####
X_valid, X_test, y_valid, y_test = train_test_split(outer_fold, outer_labels, test_size=0.5, random_state=42, stratify=outer_labels)
train_set = SignalAndTarget(inner_fold, inner_labels)
valid_set = SignalAndTarget(X_valid, y_valid)
test_set = SignalAndTarget(X_test, y_test)
if model_type == 'shallow':
model = ShallowFBCSPNet(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length,
n_filters_time=60, filter_time_length=5, n_filters_spat=40,
pool_time_length=50, pool_time_stride=15, final_conv_length='auto',
conv_nonlin=relu6, pool_mode='mean', pool_nonlin=safe_log,
split_first_layer=True, batch_norm=True, batch_norm_alpha=0.1,
drop_prob=0.1).create_network() #50 works better than 75
optimizer = optim.Adadelta(model.parameters(), lr=2.0, rho=0.9, weight_decay=0.1, eps=1e-8)
elif model_type == 'deep':
model = Deep4Net(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length,
final_conv_length='auto', n_filters_time=20, n_filters_spat=20, filter_time_length=5,
pool_time_length=3, pool_time_stride=3, n_filters_2=20, filter_length_2=5,
n_filters_3=40, filter_length_3=5, n_filters_4=1500, filter_length_4=10,
first_nonlin=leaky_relu, first_pool_mode='mean', first_pool_nonlin=safe_log, later_nonlin=leaky_relu,
later_pool_mode='mean', later_pool_nonlin=safe_log, drop_prob=0.1,
double_time_convs=False, split_first_layer=True, batch_norm=True, batch_norm_alpha=0.1,
stride_before_pool=False).create_network()
optimizer = AdamW(model.parameters(), lr=0.1, weight_decay=0)
elif model_type == 'eegnet':
model = EEGNetv4(in_chans=n_chans, n_classes=n_classes, final_conv_length='auto',
input_time_length=input_time_length, pool_mode='mean', F1=16, D=2, F2=32,
kernel_length=64, third_kernel_size=(8,4), drop_prob=0.1).create_network()
optimizer = optim.Adam(model.parameters(), lr=0.1, weight_decay=0, eps=1e-8, amsgrad=False)
if cuda:
model.cuda()
torch.backends.cudnn.deterministic = True
#model = torch.nn.DataParallel(model)
log.info("Optimized model")
model_loss_function=None
#####Setup to run the optimized model#####
optimized_model = op_exp(model, train_set, valid_set, test_set=test_set, iterator=iterator,
loss_function=best_loss, optimizer=optimizer,
model_constraint=model_constraint, monitors=monitors,
stop_criterion=stop_criterion, remember_best_column='valid_misclass',
run_after_early_stop=True, model_loss_function=model_loss_function, cuda=cuda,
data_type=data_type, subject_id=subject_id, model_type=model_type,
cropped=cropped, model_number=str(out_fold_num))
optimized_model.run()
log.info("Last 5 epochs")
log.info("\n" + str(optimized_model.epochs_df.iloc[-5:]))
writer = f"results_folder\\results\\S{subject_id}\\{data_type}_{model_type}_{str(out_fold_num)}.xlsx"
optimized_model.epochs_df.iloc[-30:].to_excel(writer)
accuracy = 1 - np.min(np.array(optimized_model.class_acc))
cv_scores.append(accuracy) # k accuracy scores for this param set.
#####Print and store fold accuracies and mean accuracy#####
print(f"Class Accuracy: {np.mean(np.array(cv_scores))}")
results_df = pd.DataFrame(dict(cv_scores=cv_scores,
cv_mean=np.mean(np.array(cv_scores))))
writer2 = f"results_folder\\results\\S{subject_id}\\{data_type}_{model_type}_cvscores.xlsx"
results_df.to_excel(writer2)
return optimized_model, np.mean(np.array(cv_scores))
def run_experiment(train_set, valid_set, test_set, model_name, optimizer_name,
init_lr, scheduler_name, use_norm_constraint, weight_decay,
schedule_weight_decay, restarts, max_epochs,
max_increase_epochs, np_th_seed):
set_random_seeds(np_th_seed, cuda=True)
#torch.backends.cudnn.benchmark = True# sometimes crashes?
if valid_set is not None:
assert max_increase_epochs is not None
assert (max_epochs is None) != (restarts is None)
if max_epochs is None:
max_epochs = np.sum(restarts)
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()
elif model_name in [
'resnet-he-uniform', 'resnet-he-normal', 'resnet-xavier-normal',
'resnet-xavier-uniform'
]:
init_name = model_name.lstrip('resnet-')
from torch.nn import init
init_fn = {
'he-uniform': lambda w: init.kaiming_uniform(w, a=0),
'he-normal': lambda w: init.kaiming_normal(w, a=0),
'xavier-uniform': lambda w: init.xavier_uniform(w, gain=1),
'xavier-normal': lambda w: init.xavier_normal(w, gain=1)
}[init_name]
model = EEGResNet(in_chans=n_chans,
n_classes=n_classes,
input_time_length=input_time_length,
final_pool_length=10,
n_first_filters=48,
conv_weight_init_fn=init_fn).create_network()
else:
raise ValueError("Unknown model name {:s}".format(model_name))
if 'resnet' not in model_name:
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]
if optimizer_name == 'adam':
optimizer = optim.Adam(model.parameters(),
weight_decay=weight_decay,
lr=init_lr)
elif optimizer_name == 'adamw':
optimizer = AdamW(model.parameters(),
weight_decay=weight_decay,
lr=init_lr)
iterator = CropsFromTrialsIterator(batch_size=60,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input,
seed=np_th_seed)
if scheduler_name is not None:
assert schedule_weight_decay == (optimizer_name == 'adamw')
if scheduler_name == 'cosine':
n_updates_per_epoch = sum(
[1 for _ in iterator.get_batches(train_set, shuffle=True)])
if restarts is None:
n_updates_per_period = n_updates_per_epoch * max_epochs
else:
n_updates_per_period = np.array(restarts) * n_updates_per_epoch
scheduler = CosineAnnealing(n_updates_per_period)
optimizer = ScheduledOptimizer(
scheduler,
optimizer,
schedule_weight_decay=schedule_weight_decay)
elif scheduler_name == 'cut_cosine':
# TODO: integrate with if clause before, now just separate
# to avoid messing with code
n_updates_per_epoch = sum(
[1 for _ in iterator.get_batches(train_set, shuffle=True)])
if restarts is None:
n_updates_per_period = n_updates_per_epoch * max_epochs
else:
n_updates_per_period = np.array(restarts) * n_updates_per_epoch
scheduler = CutCosineAnnealing(n_updates_per_period)
optimizer = ScheduledOptimizer(
scheduler,
optimizer,
schedule_weight_decay=schedule_weight_decay)
else:
raise ValueError("Unknown scheduler")
monitors = [
LossMonitor(),
MisclassMonitor(col_suffix='sample_misclass'),
CroppedTrialMisclassMonitor(input_time_length=input_time_length),
RuntimeMonitor()
]
if use_norm_constraint:
model_constraint = MaxNormDefaultConstraint()
else:
model_constraint = None
# change here this cell
loss_function = lambda preds, targets: F.nll_loss(th.mean(preds, dim=2),
targets)
if valid_set is not None:
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)
])
else:
run_after_early_stop = False
do_early_stop = False
remember_best_column = None
stop_criterion = MaxEpochs(max_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 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 main(
out_hostname, out_port, base_name, params_filename, plot_sensors,
use_out_server, adapt_model, save_data, n_updates_per_break, batch_size,
learning_rate, n_min_trials, trial_start_offset, break_start_offset,
break_stop_offset,
pred_gap,
incoming_port, load_old_data, use_new_adam_params,
input_time_length,
train_on_breaks,
min_break_samples,
min_trial_samples,
n_chans,
cuda):
setup_logging()
hostname = ''
supplier = None
sender = None
buffer = None
# load model to correct gpu id
gpu_id = th.FloatTensor(1).cuda().get_device()
def inner_device_mapping(storage, location):
return storage.cuda(gpu_id)
model_name = os.path.join(base_name, 'model.pkl')
model = th.load(model_name, map_location=inner_device_mapping)
predictor = ModelPredictor(
model, input_time_length=input_time_length, pred_gap=pred_gap,
cuda=cuda)
if adapt_model:
loss_function = log_categorical_crossentropy
model_loss_function = None
model_constraint = MaxNormDefaultConstraint()
optimizer = Adam(model.parameters(), lr=learning_rate)
n_preds_per_input = None # set later
n_classes = None # set later
trainer = BatchCntTrainer(
model, loss_function, model_loss_function, model_constraint,
optimizer, input_time_length,
n_preds_per_input, n_classes,
n_updates_per_break=n_updates_per_break,
batch_size=batch_size,
n_min_trials=n_min_trials,
trial_start_offset=trial_start_offset,
break_start_offset=break_start_offset,
break_stop_offset=break_stop_offset,
add_breaks=train_on_breaks,
min_break_samples=min_break_samples,
min_trial_samples=min_trial_samples,
cuda=cuda)
trainer.set_n_chans(n_chans)
else:
trainer = NoTrainer()
if params_filename is not None:
if params_filename == 'newest':
# sort will already sort temporally with our time string format
all_params_files = sorted(glob(os.path.join(base_name,
"*.model_params.pkl")))
assert len(all_params_files) > 0, ("Expect atleast one params file "
"if 'newest' given as argument")
params_filename = all_params_files[-1]
log.info("Loading model params from {:s}".format(params_filename))
model_params = th.load(params_filename, map_location=inner_device_mapping)
model.load_state_dict(model_params)
train_params_filename = params_filename.replace('model_params.pkl',
'trainer_params.pkl')
if os.path.isfile(train_params_filename):
if adapt_model and use_new_adam_params:
log.info("Loading trainer params from {:s}".format(
train_params_filename))
train_params = th.load(train_params_filename,
map_location=inner_device_mapping)
optimizer.load_state_dict(train_params)
elif adapt_model:
log.warn("No train/adam params found, starting optimization params "
"from scratch (model params will be loaded anyways).")
processor = StandardizeProcessor()
if adapt_model and load_old_data:
trainer.add_data_from_today(
factor_new=processor.factor_new, eps=processor.eps)
online_exp = OnlineExperiment(
supplier=supplier, buffer=buffer,
processor=processor,
predictor=predictor, trainer=trainer, sender=sender)
server = PredictionServer(
(hostname, incoming_port),
online_experiment=online_exp,
out_hostname=out_hostname, out_port=out_port,
plot_sensors=plot_sensors,
use_out_server=use_out_server, save_data=save_data,
model_base_name=base_name,
save_model_trainer_params=adapt_model)
# Compilation takes some time so initialize trainer already
# before waiting in connection in server
log.info("Starting server on port {:d}".format(incoming_port))
server.start()
log.info("Started server")
server.serve_forever()
def run_exp(data_folder, session_id, subject_id, low_cut_hz, model, cuda):
ival = [-500, 4000]
max_epochs = 1600
max_increase_epochs = 160
batch_size = 10
high_cut_hz = 38
factor_new = 1e-3
init_block_size = 1000
valid_set_fraction = .2
''' # BCIcompetition
train_filename = 'A{:02d}T.gdf'.format(subject_id)
test_filename = 'A{:02d}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)
train_cnt = train_loader.load()
test_cnt = test_loader.load()
'''
# GIGAscience
filename = 'sess{:02d}_subj{:02d}_EEG_MI.mat'.format(
session_id, subject_id)
filepath = os.path.join(data_folder, filename)
train_variable = 'EEG_MI_train'
test_variable = 'EEG_MI_test'
train_loader = GIGAscience(filepath, train_variable)
test_loader = GIGAscience(filepath, test_variable)
train_cnt = train_loader.load()
test_cnt = test_loader.load()
# Preprocessing
''' channel
['Fp1', 'Fp2', 'F7', 'F3', 'Fz', 'F4', 'F8', 'FC5', 'FC1', 'FC2', 'FC6', 'T7', 'C3', 'Cz', 'C4', 'T8', 'TP9', 'CP5',
'CP1', 'CP2', 'CP6', 'TP10', 'P7', 'P3', 'Pz', 'P4', 'P8', 'PO9', 'O1', 'Oz', 'O2', 'PO10', 'FC3', 'FC4', 'C5',
'C1', 'C2', 'C6', 'CP3', 'CPz', 'CP4', 'P1', 'P2', 'POz', 'FT9', 'FTT9h', 'TTP7h', 'TP7', 'TPP9h', 'FT10',
'FTT10h', 'TPP8h', 'TP8', 'TPP10h', 'F9', 'F10', 'AF7', 'AF3', 'AF4', 'AF8', 'PO3', 'PO4']
'''
train_cnt = train_cnt.pick_channels([
'FC5', 'FC3', 'FC1', 'Fz', 'FC2', 'FC4', 'FC6', 'C5', 'C3', 'C1', 'Cz',
'C2', 'C4', 'C6', 'CP5', 'CP3', 'CP1', 'CPz', 'CP2', 'CP4', 'CP6', 'Pz'
])
train_cnt, train_cnt.info['events'] = train_cnt.copy().resample(
250, npad='auto', events=train_cnt.info['events'])
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,
high_cut_hz,
train_cnt.info['sfreq'],
filt_order=3,
axis=1), train_cnt)
train_cnt = mne_apply(
lambda a: exponential_running_standardize(a.T,
factor_new=factor_new,
init_block_size=
init_block_size,
eps=1e-4).T, train_cnt)
test_cnt = test_cnt.pick_channels([
'FC5', 'FC3', 'FC1', 'Fz', 'FC2', 'FC4', 'FC6', 'C5', 'C3', 'C1', 'Cz',
'C2', 'C4', 'C6', 'CP5', 'CP3', 'CP1', 'CPz', 'CP2', 'CP4', 'CP6', 'Pz'
])
test_cnt, test_cnt.info['events'] = test_cnt.copy().resample(
250, npad='auto', events=test_cnt.info['events'])
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,
high_cut_hz,
test_cnt.info['sfreq'],
filt_order=3,
axis=1), test_cnt)
test_cnt = mne_apply(
lambda a: exponential_running_standardize(a.T,
factor_new=factor_new,
init_block_size=
init_block_size,
eps=1e-4).T, test_cnt)
marker_def = OrderedDict([('Right Hand', [1]), ('Left Hand', [2])])
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=1 -
valid_set_fraction)
set_random_seeds(seed=20190706, cuda=cuda)
n_classes = 2
n_chans = int(train_set.X.shape[1])
input_time_length = train_set.X.shape[2]
if model == 'shallow':
model = ShallowFBCSPNet(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto').create_network()
elif model == 'deep':
model = Deep4Net(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length='auto').create_network()
if cuda:
model.cuda()
log.info("Model: \n{:s}".format(str(model)))
optimizer = optim.Adam(model.parameters())
iterator = BalancedBatchSizeIterator(batch_size=batch_size)
stop_criterion = Or([
MaxEpochs(max_epochs),
NoDecrease('valid_misclass', max_increase_epochs)
])
monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()]
model_constraint = MaxNormDefaultConstraint()
exp = Experiment(model,
train_set,
valid_set,
test_set,
iterator=iterator,
loss_function=F.nll_loss,
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
def run_exp(data_folder, subject_id, low_cut_hz, model, cuda):
ival = [-500, 4000]
input_time_length = 1000
max_epochs = 800
max_increase_epochs = 80
batch_size = 60
high_cut_hz = 38
factor_new = 1e-3
init_block_size = 1000
valid_set_fraction = 0.2
train_filename = 'A{:02d}T.gdf'.format(subject_id)
test_filename = 'A{:02d}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)
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,
high_cut_hz,
train_cnt.info['sfreq'],
filt_order=3,
axis=1), train_cnt)
train_cnt = mne_apply(
lambda a: exponential_running_standardize(a.T,
factor_new=factor_new,
init_block_size=
init_block_size,
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,
high_cut_hz,
test_cnt.info['sfreq'],
filt_order=3,
axis=1), test_cnt)
test_cnt = mne_apply(
lambda a: exponential_running_standardize(a.T,
factor_new=factor_new,
init_block_size=
init_block_size,
eps=1e-4).T, test_cnt)
marker_def = OrderedDict([('Left Hand', [1]), (
'Right Hand',
[2],
), ('Foot', [3]), ('Tongue', [4])])
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=1 -
valid_set_fraction)
set_random_seeds(seed=20190706, cuda=cuda)
n_classes = 4
n_chans = int(train_set.X.shape[1])
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=batch_size,
input_time_length=input_time_length,
n_preds_per_input=n_preds_per_input)
stop_criterion = Or([
MaxEpochs(max_epochs),
NoDecrease('valid_misclass', max_increase_epochs)
])
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
def 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,
cuda=True):
import torch.backends.cudnn as cudnn
cudnn.benchmark = True
if optimizer == 'adam':
assert merge_train_valid == False
else:
assert optimizer == 'adamw'
assert merge_train_valid == True
preproc_functions = create_preproc_functions(
sec_to_cut_at_start=sec_to_cut_at_start,
sec_to_cut_at_end=sec_to_cut_at_end,
duration_recording_mins=duration_recording_mins,
max_abs_val=max_abs_val,
clip_before_resample=clip_before_resample,
sampling_freq=sampling_freq,
divisor=divisor)
dataset = DiagnosisSet(n_recordings=n_recordings,
max_recording_mins=max_recording_mins,
preproc_functions=preproc_functions,
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 = create_preproc_functions(
sec_to_cut_at_start=sec_to_cut_at_start,
sec_to_cut_at_end=sec_to_cut_at_end,
duration_recording_mins=test_recording_mins,
max_abs_val=max_abs_val,
clip_before_resample=clip_before_resample,
sampling_freq=sampling_freq,
divisor=divisor)
test_dataset = DiagnosisSet(n_recordings=n_recordings,
max_recording_mins=None,
preproc_functions=test_preproc_functions,
train_or_eval='eval',
sensor_types=sensor_types)
if not only_return_exp:
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)
else:
splitter = TrainValidSplitter(n_folds,
i_valid_fold=i_test_fold,
shuffle=shuffle)
if not only_return_exp:
if not test_on_eval:
train_set, valid_set, test_set = splitter.split(X, y)
else:
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
if merge_train_valid:
train_set = concatenate_sets([train_set, valid_set])
# just reduce valid for faster computations
valid_set.X = valid_set.X[:8]
valid_set.y = valid_set.y[:8]
# np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/train_X.npy', train_set.X)
# np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/train_y.npy', train_set.y)
# np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/valid_X.npy', valid_set.X)
# np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/valid_y.npy', valid_set.y)
# np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/test_X.npy', test_set.X)
# np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/test_y.npy', test_set.y)
else:
train_set = None
valid_set = None
test_set = None
log.info("Model:\n{:s}".format(str(model)))
if cuda:
model.cuda()
model.eval()
in_chans = 21
# 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,
seed=np_th_seed)
assert optimizer in ['adam', 'adamw'], ("Expect optimizer to be either "
"adam or adamw")
schedule_weight_decay = optimizer == 'adamw'
if optimizer == 'adam':
optim_class = optim.Adam
assert schedule_weight_decay == False
assert merge_train_valid == False
else:
optim_class = AdamW
assert schedule_weight_decay == True
assert merge_train_valid == True
optimizer = optim_class(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
if scheduler is not None:
assert scheduler == 'cosine'
n_updates_per_epoch = sum(
[1 for _ in iterator.get_batches(train_set, shuffle=True)])
# Adapt if you have a different number of epochs
n_updates_per_period = n_updates_per_epoch * max_epochs
scheduler = CosineAnnealing(n_updates_per_period)
optimizer = ScheduledOptimizer(
scheduler, optimizer, schedule_weight_decay=schedule_weight_decay)
loss_function = nll_loss_on_mean
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)
loggers = [Printer(), TensorboardWriter(log_dir)]
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,
loggers=loggers)
if not only_return_exp:
# Until first stop
exp.setup_training()
exp.monitor_epoch(exp.datasets)
exp.log_epoch()
exp.rememberer.remember_epoch(exp.epochs_df, exp.model, exp.optimizer)
exp.iterator.reset_rng()
while not exp.stop_criterion.should_stop(exp.epochs_df):
if (time.time() - start_time) > time_cut_off_sec:
log.info(
"Ran out of time after {:.2f} sec.".format(time.time() -
start_time))
return exp
log.info("Still in time after {:.2f} sec.".format(time.time() -
start_time))
exp.run_one_epoch(exp.datasets, remember_best=True)
if (time.time() - start_time) > time_cut_off_sec:
log.info("Ran out of time after {:.2f} sec.".format(time.time() -
start_time))
return exp
if not merge_train_valid:
exp.setup_after_stop_training()
# Run until second stop
datasets = exp.datasets
datasets['train'] = concatenate_sets(
[datasets['train'], datasets['valid']])
exp.monitor_epoch(datasets)
exp.log_epoch()
exp.iterator.reset_rng()
while not exp.stop_criterion.should_stop(exp.epochs_df):
if (time.time() - start_time) > time_cut_off_sec:
log.info("Ran out of time after {:.2f} sec.".format(
time.time() - start_time))
return exp
log.info("Still in time after {:.2f} sec.".format(time.time() -
start_time))
exp.run_one_epoch(datasets, remember_best=False)
else:
exp.dataset = dataset
exp.splitter = splitter
if test_on_eval:
exp.test_dataset = test_dataset
return exp
def setup_exp(
train_folder,
n_recordings,
n_chans,
model_name,
n_start_chans,
n_chan_factor,
input_time_length,
final_conv_length,
model_constraint,
stride_before_pool,
init_lr,
batch_size,
max_epochs,
cuda,
num_workers,
task,
weight_decay,
n_folds,
shuffle_folds,
lazy_loading,
eval_folder,
result_folder,
run_on_normals,
run_on_abnormals,
seed,
l2_decay,
gradient_clip,
):
info_msg = "using {}, {}".format(
os.environ["SLURM_JOB_PARTITION"],
os.environ["SLURMD_NODENAME"],
)
info_msg += ", gpu {}".format(os.environ["CUDA_VISIBLE_DEVICES"])
logging.info(info_msg)
logging.info("Targets for this task: <{}>".format(task))
import torch.backends.cudnn as cudnn
cudnn.benchmark = True
if task == "age":
loss_function = mse_loss_on_mean
remember_best_column = "valid_rmse"
n_classes = 1
else:
loss_function = nll_loss_on_mean
remember_best_column = "valid_misclass"
n_classes = 2
if model_constraint is not None:
assert model_constraint == 'defaultnorm'
model_constraint = MaxNormDefaultConstraint()
stop_criterion = MaxEpochs(max_epochs)
set_random_seeds(seed=seed, cuda=cuda)
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 == 'eegnet':
model = EEGNetv4(n_chans,
n_classes,
input_time_length=input_time_length,
final_conv_length=final_conv_length).create_network()
elif model_name == "tcn":
assert task != "age", "what to change to do regression with tcn?!"
model = TemporalConvNet(input_size=n_chans,
output_size=n_classes,
context_size=0,
num_channels=55,
num_levels=5,
kernel_size=16,
dropout=0.05270154233150525,
skip_mode=None,
use_context=0,
lasso_selection=0.0,
rnn_normalization=None)
else:
assert False, "unknown model name {:s}".format(model_name)
if task == "age":
# remove softmax layer, set n_classes to 1
model.n_classes = 1
new_model = nn.Sequential()
for name, module_ in model.named_children():
if name == "softmax":
continue
new_model.add_module(name, module_)
model = new_model
# maybe check if this works and wait / re-try after some time?
# in case of all cuda devices are busy
if cuda:
model.cuda()
if model_name != "tcn":
to_dense_prediction_model(model)
logging.info("Model:\n{:s}".format(str(model)))
test_input = np_to_var(
np.ones((2, n_chans, input_time_length, 1), dtype=np.float32))
if list(model.parameters())[0].is_cuda:
test_input = test_input.cuda()
out = model(test_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
if eval_folder is None:
logging.info("will do validation")
if lazy_loading:
logging.info(
"using lazy loading to load {} recs".format(n_recordings))
dataset = TuhLazy(train_folder,
target=task,
n_recordings=n_recordings)
else:
logging.info("using traditional loading to load {} recs".format(
n_recordings))
dataset = Tuh(train_folder, n_recordings=n_recordings, target=task)
assert not (run_on_normals and run_on_abnormals), (
"decide whether to run on normal or abnormal subjects")
# only run on normal subjects
if run_on_normals:
ids = [
i for i in range(len(dataset)) if dataset.pathologicals[i] == 0
] # 0 is non-pathological
dataset = TuhSubset(dataset, ids)
logging.info("only using {} normal subjects".format(len(dataset)))
if run_on_abnormals:
ids = [
i for i in range(len(dataset)) if dataset.pathologicals[i] == 1
] # 1 is pathological
dataset = TuhSubset(dataset, ids)
logging.info("only using {} abnormal subjects".format(
len(dataset)))
indices = np.arange(len(dataset))
kf = KFold(n_splits=n_folds, shuffle=shuffle_folds)
for i, (train_ind, test_ind) in enumerate(kf.split(indices)):
assert len(np.intersect1d(
train_ind, test_ind)) == 0, ("train and test set overlap!")
# seed is in range of number of folds and was set by submit script
if i == seed:
break
if lazy_loading:
test_subset = TuhLazySubset(dataset, test_ind)
train_subset = TuhLazySubset(dataset, train_ind)
else:
test_subset = TuhSubset(dataset, test_ind)
train_subset = TuhSubset(dataset, train_ind)
else:
logging.info("will do final evaluation")
if lazy_loading:
train_subset = TuhLazy(train_folder, target=task)
test_subset = TuhLazy(eval_folder, target=task)
else:
train_subset = Tuh(train_folder, target=task)
test_subset = Tuh(eval_folder, target=task)
# remove rec:
# train/abnormal/01_tcp_ar/081/00008184/s001_2011_09_21/00008184_s001_t001
# since it contains no crop without outliers (channels A1, A2 broken)
subjects = [f.split("/")[-3] for f in train_subset.file_paths]
if "00008184" in subjects:
bad_id = subjects.index("00008184")
train_subset = remove_file_from_dataset(
train_subset,
file_id=bad_id,
file=("train/abnormal/01_tcp_ar/081/00008184/s001_2011_09_21/"
"00008184_s001_t001"))
subjects = [f.split("/")[-3] for f in test_subset.file_paths]
if "00008184" in subjects:
bad_id = subjects.index("00008184")
test_subset = remove_file_from_dataset(
test_subset,
file_id=bad_id,
file=("train/abnormal/01_tcp_ar/081/00008184/s001_2011_09_21/"
"00008184_s001_t001"))
if task == "age":
# standardize ages based on train set
y_train = train_subset.y
y_train_mean = np.mean(y_train)
y_train_std = np.std(y_train)
train_subset.y = (y_train - y_train_mean) / y_train_std
y_test = test_subset.y
test_subset.y = (y_test - y_train_mean) / y_train_std
if lazy_loading:
iterator = LazyCropsFromTrialsIterator(
input_time_length,
n_preds_per_input,
batch_size,
seed=seed,
num_workers=num_workers,
reset_rng_after_each_batch=False,
check_preds_smaller_trial_len=False) # True!
else:
iterator = CropsFromTrialsIterator(batch_size, input_time_length,
n_preds_per_input, seed)
monitors = []
monitors.append(LossMonitor())
monitors.append(RAMMonitor())
monitors.append(RuntimeMonitor())
if task == "age":
monitors.append(
RMSEMonitor(input_time_length,
n_preds_per_input,
mean=y_train_mean,
std=y_train_std))
else:
monitors.append(
CroppedDiagnosisMonitor(input_time_length, n_preds_per_input))
monitors.append(LazyMisclassMonitor(col_suffix='sample_misclass'))
if lazy_loading:
n_updates_per_epoch = len(
iterator.get_batches(train_subset, shuffle=False))
else:
n_updates_per_epoch = sum(
[1 for _ in iterator.get_batches(train_subset, shuffle=False)])
n_updates_per_period = n_updates_per_epoch * max_epochs
logging.info("there are {} updates per epoch".format(n_updates_per_epoch))
if model_name == "tcn":
adamw = ExtendedAdam(model.parameters(),
lr=init_lr,
weight_decay=weight_decay,
l2_decay=l2_decay,
gradient_clip=gradient_clip)
else:
adamw = AdamWWithTracking(model.parameters(),
init_lr,
weight_decay=weight_decay)
scheduler = CosineAnnealing(n_updates_per_period)
optimizer = ScheduledOptimizer(scheduler,
adamw,
schedule_weight_decay=True)
exp = Experiment(model=model,
train_set=train_subset,
valid_set=None,
test_set=test_subset,
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=False,
batch_modifier=None,
cuda=cuda,
do_early_stop=False,
reset_after_second_run=False)
return exp