def setup_after_stop_training(self):
"""
Setup training after first stop.
Resets parameters to best parameters and updates stop criterion.
"""
# also remember old monitor chans, will be put back into
# monitor chans after experiment finished
self.before_stop_df = deepcopy(self.epochs_df)
self.rememberer.reset_to_best_model(self.epochs_df, self.model,
self.optimizer)
loss_to_reach = float(self.epochs_df['train_loss'].iloc[-1])
self.stop_criterion = Or(stop_criteria=[
MaxEpochs(max_epochs=self.rememberer.best_epoch * 2),
ColumnBelow(column_name='valid_loss', target_value=loss_to_reach)
])
log.info("Train loss to reach {:.5f}".format(loss_to_reach))
# def exp_lr_scheduler(optimizer, global_step, init_lr, decay_steps, decay_rate, lr_clip, staircase=True):
# if staircase:
# lr = init_lr * decay_rate
# else:
# lr = init_lr * decay_rate**(global_step / decay_steps)
# lr = max(lr, lr_clip)
# #print(lr)
# if global_step % decay_steps == 0:
# print('LR is set to {}'.format(lr))
# lr=0.1
# for param_group in optimizer.param_groups:
# param_group['lr'] = lr
# return optimizer
# def adjust_lr(optimizer, epoch):
# lr = init_lr * (0.1 ** (epoch // 20))
# for param_group in optimizer.param_groups:
# param_group['lr'] = lr
# return optimizer
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 train_model(self, train_set, val_set, test_set, save_model):
"""
:param train_set: EEG data (n_trials*n_channels*n_samples)
:param val_set: EEG data (n_trials*n_channels*n_samples)
:param test_set: EEG data (n_trials*n_channels*n_samples) - can be None when training on inner-fold
:param save_model: Boolean: True if trained model is to be saved
:return: Accuracy and loss scores for the model trained with a given set of hyper-parameters
"""
predictions = None
model = None
model = self.call_model()
set_random_seeds(seed=20190629, cuda=self.cuda)
if self.cuda:
model.cuda()
torch.backends.cudnn.deterministic = True
log.info("%s model: ".format(str(model)))
optimizer = optim.Adam(model.parameters(),
lr=self.learning_rate,
weight_decay=0,
eps=1e-8,
amsgrad=False)
stop_criterion = Or([
MaxEpochs(self.epochs),
NoDecrease('valid_misclass', self.max_increase_epochs)
])
model_loss_function = None
#####Setup to run the selected model#####
model_test = Experiment(model,
train_set,
val_set,
test_set=test_set,
iterator=self.iterator,
loss_function=self.loss,
optimizer=optimizer,
model_constraint=self.model_constraint,
monitors=self.monitors,
stop_criterion=stop_criterion,
remember_best_column='valid_misclass',
run_after_early_stop=True,
model_loss_function=model_loss_function,
cuda=self.cuda,
data_type=self.data_type,
model_type=self.model_type,
subject_id=self.subject,
model_number=str(self.model_number),
save_model=save_model)
model_test.run()
model_acc = model_test.epochs_df['valid_misclass'].astype('float')
model_loss = model_test.epochs_df['valid_loss'].astype('float')
current_val_acc = 1 - current_acc(model_acc)
current_val_loss = current_loss(model_loss)
test_accuracy = None
if test_set is not None:
test_accuracy = round(
(1 - model_test.epochs_df['test_misclass'].min()) * 100, 3)
predictions = model_test.predictions
probabilities = model_test.probabilites
return current_val_acc, current_val_loss, test_accuracy, model_test, predictions, probabilities
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
def train_model(self,
model,
dataset,
state=None,
final_evaluation=False,
ensemble=False):
if type(model) == list:
model = AveragingEnsemble(model)
if self.cuda:
for mod in model.models:
mod.cuda()
if self.cuda:
torch.cuda.empty_cache()
if final_evaluation:
self.stop_criterion = Or([
MaxEpochs(global_vars.get('final_max_epochs')),
NoIncreaseDecrease(
f'valid_{global_vars.get("nn_objective")}',
global_vars.get('final_max_increase_epochs'),
oper=get_oper_by_loss_function(self.loss_function))
])
if global_vars.get('cropping'):
self.set_cropping_for_model(model)
self.epochs_df = pd.DataFrame()
if global_vars.get('do_early_stop') or global_vars.get(
'remember_best'):
self.rememberer = RememberBest(
f"valid_{global_vars.get('nn_objective')}",
oper=get_oper_by_loss_function(self.loss_function,
equals=True))
self.optimizer = optim.Adam(model.parameters())
if self.cuda:
assert torch.cuda.is_available(), "Cuda not available"
if torch.cuda.device_count() > 1 and global_vars.get(
'parallel_gpu'):
model.cuda()
with torch.cuda.device(0):
model = nn.DataParallel(
model.cuda(),
device_ids=[
int(s)
for s in global_vars.get('gpu_select').split(',')
])
else:
model.cuda()
try:
if global_vars.get('inherit_weights_normal') and state is not None:
current_state = model.state_dict()
for k, v in state.items():
if k in current_state and current_state[k].shape == v.shape:
current_state.update({k: v})
model.load_state_dict(current_state)
except Exception as e:
print(f'failed weight inheritance\n,'
f'state dict: {state.keys()}\n'
f'current model state: {model.state_dict().keys()}')
print('load state dict failed. Exception message: %s' % (str(e)))
pdb.set_trace()
self.monitor_epoch(dataset, model)
if global_vars.get('log_epochs'):
self.log_epoch()
if global_vars.get('remember_best'):
self.rememberer.remember_epoch(self.epochs_df, model,
self.optimizer)
self.iterator.reset_rng()
start = time.time()
num_epochs = self.run_until_stop(model, dataset)
self.setup_after_stop_training(model, final_evaluation)
if final_evaluation:
dataset_train_backup = deepcopy(dataset['train'])
if ensemble:
self.run_one_epoch(dataset, model)
self.rememberer.remember_epoch(self.epochs_df,
model,
self.optimizer,
force=ensemble)
num_epochs += 1
else:
dataset['train'] = concatenate_sets(
[dataset['train'], dataset['valid']])
num_epochs += self.run_until_stop(model, dataset)
self.rememberer.reset_to_best_model(self.epochs_df, model,
self.optimizer)
dataset['train'] = dataset_train_backup
end = time.time()
self.final_time = end - start
self.num_epochs = num_epochs
return model
class NN_Trainer:
def __init__(self, iterator, loss_function, stop_criterion, monitors):
global model_train_times
model_train_times = []
self.iterator = iterator
self.loss_function = loss_function
self.optimizer = None
self.rememberer = None
self.loggers = [Printer()]
self.stop_criterion = stop_criterion
self.monitors = monitors
self.cuda = global_vars.get('cuda')
self.loggers = [Printer()]
self.epochs_df = None
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 get_n_preds_per_input(self, model):
dummy_input = self.get_dummy_input()
if global_vars.get('cuda'):
model.cuda()
dummy_input = dummy_input.cuda()
out = model(dummy_input)
n_preds_per_input = out.cpu().data.numpy().shape[2]
return n_preds_per_input
def train_ensemble(self, models, dataset, final_evaluation=False):
# for i in range(len(models)):
# models[i] = self.train_model(models[i], dataset, final_evaluation=final_evaluation)
models = [
nn.Sequential(
*list(model.children())[:global_vars.get('num_layers') + 1])
for model in models
] # remove the final softmax layer from each model
avg_model = AveragingEnsemble(models)
return self.train_model(avg_model,
dataset,
final_evaluation=final_evaluation)
def train_model(self,
model,
dataset,
state=None,
final_evaluation=False,
ensemble=False):
if type(model) == list:
model = AveragingEnsemble(model)
if self.cuda:
for mod in model.models:
mod.cuda()
if self.cuda:
torch.cuda.empty_cache()
if final_evaluation:
self.stop_criterion = Or([
MaxEpochs(global_vars.get('final_max_epochs')),
NoIncreaseDecrease(
f'valid_{global_vars.get("nn_objective")}',
global_vars.get('final_max_increase_epochs'),
oper=get_oper_by_loss_function(self.loss_function))
])
if global_vars.get('cropping'):
self.set_cropping_for_model(model)
self.epochs_df = pd.DataFrame()
if global_vars.get('do_early_stop') or global_vars.get(
'remember_best'):
self.rememberer = RememberBest(
f"valid_{global_vars.get('nn_objective')}",
oper=get_oper_by_loss_function(self.loss_function,
equals=True))
self.optimizer = optim.Adam(model.parameters())
if self.cuda:
assert torch.cuda.is_available(), "Cuda not available"
if torch.cuda.device_count() > 1 and global_vars.get(
'parallel_gpu'):
model.cuda()
with torch.cuda.device(0):
model = nn.DataParallel(
model.cuda(),
device_ids=[
int(s)
for s in global_vars.get('gpu_select').split(',')
])
else:
model.cuda()
try:
if global_vars.get('inherit_weights_normal') and state is not None:
current_state = model.state_dict()
for k, v in state.items():
if k in current_state and current_state[k].shape == v.shape:
current_state.update({k: v})
model.load_state_dict(current_state)
except Exception as e:
print(f'failed weight inheritance\n,'
f'state dict: {state.keys()}\n'
f'current model state: {model.state_dict().keys()}')
print('load state dict failed. Exception message: %s' % (str(e)))
pdb.set_trace()
self.monitor_epoch(dataset, model)
if global_vars.get('log_epochs'):
self.log_epoch()
if global_vars.get('remember_best'):
self.rememberer.remember_epoch(self.epochs_df, model,
self.optimizer)
self.iterator.reset_rng()
start = time.time()
num_epochs = self.run_until_stop(model, dataset)
self.setup_after_stop_training(model, final_evaluation)
if final_evaluation:
dataset_train_backup = deepcopy(dataset['train'])
if ensemble:
self.run_one_epoch(dataset, model)
self.rememberer.remember_epoch(self.epochs_df,
model,
self.optimizer,
force=ensemble)
num_epochs += 1
else:
dataset['train'] = concatenate_sets(
[dataset['train'], dataset['valid']])
num_epochs += self.run_until_stop(model, dataset)
self.rememberer.reset_to_best_model(self.epochs_df, model,
self.optimizer)
dataset['train'] = dataset_train_backup
end = time.time()
self.final_time = end - start
self.num_epochs = num_epochs
return model
def train_with_skorch(self, model, dataset):
if dataset['train'].X.ndim == 3:
dataX = deepcopy(dataset['train'].X)
dataX = dataX[:, :, :, None]
else:
dataX = dataset['train'].X
net = NeuralNetClassifier(
model,
max_epochs=global_vars.get('max_epochs'),
lr=1e-3,
# Shuffle training data on each epoch
iterator_train__shuffle=True,
)
net.fit(dataX, dataset['train'].y)
return net
def train_and_evaluate_model(self,
model,
dataset,
state=None,
final_evaluation=False,
ensemble=False):
if global_vars.get('train_with_skorch'):
model = self.train_with_skorch(model, dataset)
pass
else:
self.train_model(model, dataset, state, final_evaluation, ensemble)
evaluations = {}
for evaluation_metric in global_vars.get('evaluation_metrics'):
evaluations[evaluation_metric] = {
'train': self.epochs_df.iloc[-1][f"train_{evaluation_metric}"],
'valid': self.epochs_df.iloc[-1][f"valid_{evaluation_metric}"],
'test': self.epochs_df.iloc[-1][f"test_{evaluation_metric}"]
}
if self.cuda:
torch.cuda.empty_cache()
if global_vars.get('delete_finalized_models'):
if global_vars.get('grid_as_ensemble'):
model_stats = {}
for param_idx, param in enumerate(
list(model.pytorch_layers['averaging_layer'].
parameters())):
for inner_idx, inner_param in enumerate(param[0]):
model_stats[
f'avg_weights {(param_idx, inner_idx)}'] = float(
inner_param)
del model
model = model_stats
else:
del model
model = None
return self.final_time, evaluations, model, self.rememberer.model_state_dict, self.num_epochs
def setup_after_stop_training(self, model, final_evaluation):
self.rememberer.reset_to_best_model(self.epochs_df, model,
self.optimizer)
if final_evaluation:
loss_to_reach = float(self.epochs_df['train_loss'].iloc[-1])
self.stop_criterion = Or(stop_criteria=[
MaxEpochs(max_epochs=global_vars.get('final_max_epochs')),
ColumnBelow(column_name='valid_loss',
target_value=loss_to_reach)
])
def run_until_stop(self, model, single_subj_dataset):
num_epochs = 0
while not self.stop_criterion.should_stop(self.epochs_df):
self.run_one_epoch(single_subj_dataset, model)
num_epochs += 1
return num_epochs
def run_one_epoch(self, datasets, model):
model.train()
batch_generator = self.iterator.get_batches(datasets['train'],
shuffle=True)
for inputs, targets in batch_generator:
input_vars = np_to_var(inputs,
pin_memory=global_vars.get('pin_memory'))
target_vars = np_to_var(targets,
pin_memory=global_vars.get('pin_memory'))
if self.cuda:
with torch.cuda.device(0):
input_vars = input_vars.cuda()
target_vars = target_vars.cuda()
self.optimizer.zero_grad()
outputs = model(input_vars)
if self.loss_function == F.mse_loss:
target_vars = target_vars.float()
loss = self.loss_function(outputs.squeeze(), target_vars)
loss.backward()
self.optimizer.step()
self.monitor_epoch(datasets, model)
if global_vars.get('log_epochs'):
self.log_epoch()
if global_vars.get('remember_best'):
self.rememberer.remember_epoch(self.epochs_df, model,
self.optimizer)
def monitor_epoch(self, datasets, model):
result_dicts_per_monitor = OrderedDict()
for m in self.monitors:
result_dicts_per_monitor[m] = OrderedDict()
for m in self.monitors:
result_dict = m.monitor_epoch()
if result_dict is not None:
result_dicts_per_monitor[m].update(result_dict)
raws = {}
targets = {}
for setname in datasets:
assert setname in ['train', 'valid', 'test']
dataset = datasets[setname]
all_preds = []
all_losses = []
all_batch_sizes = []
all_targets = []
for batch in self.iterator.get_batches(dataset, shuffle=False):
input_vars = batch[0]
target_vars = batch[1]
preds, loss = self.eval_on_batch(input_vars, target_vars,
model)
all_preds.append(preds)
all_losses.append(loss)
all_batch_sizes.append(len(input_vars))
all_targets.append(target_vars)
for m in self.monitors:
result_dict = m.monitor_set(setname, all_preds, all_losses,
all_batch_sizes, all_targets,
dataset)
if result_dict is not None:
result_dicts_per_monitor[m].update(result_dict)
if global_vars.get('ensemble_iterations'):
raws.update({
f'{setname}_raw':
list(itertools.chain.from_iterable(all_preds))
})
targets.update({
f'{setname}_target':
list(itertools.chain.from_iterable(all_targets))
})
row_dict = OrderedDict()
if global_vars.get('ensemble_iterations'):
row_dict.update(raws)
row_dict.update(targets)
for m in self.monitors:
row_dict.update(result_dicts_per_monitor[m])
self.epochs_df = self.epochs_df.append(row_dict, ignore_index=True)
assert set(self.epochs_df.columns) == set(row_dict.keys())
self.epochs_df = self.epochs_df[list(row_dict.keys())]
def eval_on_batch(self, inputs, targets, model):
"""
Evaluate given inputs and targets.
Parameters
----------
inputs: `torch.autograd.Variable`
targets: `torch.autograd.Variable`
Returns
-------
predictions: `torch.autograd.Variable`
loss: `torch.autograd.Variable`
"""
model.eval()
with torch.no_grad():
input_vars = np_to_var(inputs,
pin_memory=global_vars.get('pin_memory'))
target_vars = np_to_var(targets,
pin_memory=global_vars.get('pin_memory'))
if self.cuda:
with torch.cuda.device(0):
input_vars = input_vars.cuda()
target_vars = target_vars.cuda()
outputs = model(input_vars)
if self.loss_function == F.mse_loss:
target_vars = target_vars.float()
loss = self.loss_function(outputs.squeeze(), target_vars)
if hasattr(outputs, 'cpu'):
outputs = outputs.cpu().data.numpy()
else:
# assume it is iterable
outputs = [o.cpu().data.numpy() for o in outputs]
loss = loss.cpu().data.numpy()
return outputs, loss
def log_epoch(self):
"""
Print monitoring values for this epoch.
"""
for logger in self.loggers:
logger.log_epoch(self.epochs_df)
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(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_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
batch_size = 60
max_epochs = 20000
max_increase_epochs = 360
model = ShallowFBCSPNet(in_chan,
db.n_classes,
input_time_length=time_steps,
final_conv_length="auto").create_network()
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,
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 train_model(self, train_set_1, val_set_1, test_set_1, train_set_2,
val_set_2, test_set_2, save_model):
"""
:param train_set_1: (np.array) n_trials*n_channels*n_samples
:param val_set_1: (np.array) n_trials*n_channels*n_samples
:param test_set_1: (np.array) n_trials*n_channels*n_samples - can be None when training on inner-fold
:param train_set_2: (np.array) n_trials*n_channels*n_samples
:param val_set_2: (np.array) n_trials*n_channels*n_samples
:param test_set_2: (np.array) n_trials*n_channels*n_samples - can be None when training on inner-fold
:param save_model: (Bool) True if trained model is to be saved
:return: Accuracy and loss scores for the model trained with a given set of hyper-parameters
"""
model = self.call_model()
predictions = None
set_random_seeds(seed=20190629, cuda=self.cuda)
if self.cuda:
model.cuda()
torch.backends.cudnn.deterministic = True
model = torch.nn.DataParallel(model)
log.info(f"Cuda in use")
log.info("%s model: ".format(str(model)))
optimizer = optim.Adam(model.parameters(),
lr=self.learning_rate,
weight_decay=0.01,
eps=1e-8,
amsgrad=False)
stop_criterion = Or([
MaxEpochs(self.epochs),
NoDecrease('valid_loss', self.max_increase_epochs)
])
model_loss_function = None
#####Setup to run the selected model#####
model_test = Experiment(model,
train_set_1,
val_set_1,
train_set_2,
val_set_2,
test_set_1=test_set_1,
test_set_2=test_set_2,
iterator=self.iterator,
loss_function=self.loss,
optimizer=optimizer,
lr_scheduler=self.lr_scheduler(
optimizer,
step_size=self.lr_step,
gamma=self.lr_gamma),
model_constraint=self.model_constraint,
monitors=self.monitors,
stop_criterion=stop_criterion,
remember_best_column='valid_misclass',
run_after_early_stop=True,
model_loss_function=model_loss_function,
cuda=self.cuda,
save_file=self.model_save_path,
tag=self.tag,
save_model=save_model)
model_test.run()
model_acc = model_test.epochs_df['valid_misclass'].astype('float')
model_loss = model_test.epochs_df['valid_loss'].astype('float')
current_val_acc = 1 - current_acc(model_acc)
current_val_loss = current_loss(model_loss)
test_accuracy = None
if train_set_1 is not None and test_set_2 is not None:
val_metric_index = self.get_model_index(model_test.epochs_df)
test_accuracy = round(
(1 -
model_test.epochs_df['test_misclass'].iloc[val_metric_index])
* 100, 3)
predictions = model_test.model_predictions
probabilities = model_test.model_probabilities
return current_val_acc, current_val_loss, test_accuracy, model_test, predictions, probabilities