-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
282 lines (240 loc) · 14.6 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
import warnings
import os
import torch
from torch import nn
from torch.utils.data import DataLoader
from model import LearningRateScheduler
def training(parameter, model, train_data_set, test_data_set, n_classes, num_workers=1):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)
# log with WandB
if parameter['wandb_log']:
import wandb
# get rid of parameter that should not be logged
wandb_dict = {k: parameter[k] for k in parameter.keys() - ['wandb_log', 'compile_model', 'wandb_proj',
'data_set', 'wandb_name']}
if parameter['wandb_name']:
wandb.init(project=parameter['wandb_proj'], config=wandb_dict, reinit=False, name=parameter['wandb_name'])
else:
wandb.init(project=parameter['wandb_proj'], config=wandb_dict, reinit=False)
if parameter['wandb_watch']:
wandb.watch(models=model, log='all', log_freq=25)
# Loss function
criterion = nn.CrossEntropyLoss(label_smoothing=parameter['label_smoothing']).to(device)
optimizer = model.configure_optimizers(weight_decay=parameter['weight_decay'],
weight_decay_cls_head=parameter['weight_decay_cls_head'],
learning_rate=parameter['lr'], betas=parameter['betas'], device_type=device)
'''
alternative simple optimizer
optimizer = torch.optim.AdamW(params=model.parameters(), lr=parameter['lr'], betas=(0.9, 0.95),
weight_decay=parameter['weight_decay'])
'''
learning_rate_scheduler = LearningRateScheduler(warmup_iters=parameter['lr_warm_up_iters'],
learning_rate=parameter['lr'],
lr_decay_iters=parameter['num_epochs'], min_lr=parameter['lr']*1/10)
# compile_model
if parameter['compile_model']:
# torch.compile requires PyTorch >=2.0
if int(torch.__version__[0]) >= 2:
model = torch.compile(model)
print('Model is compiled')
else:
warnings.warn('Compile model requires PyTorch >= 2.0, model is NOT compiled!')
# --------------------------------------------------------------------------------------------------------------
# Training Loop
if not train_data_set.dict_channels == test_data_set.dict_channels:
raise ValueError('Channel index between train and test is not consistence')
train_generator = torch.utils.data.DataLoader(train_data_set, batch_size=parameter['batch_size'], shuffle=True,
drop_last=True,
collate_fn=train_data_set.my_collate, num_workers=num_workers)
test_generator = torch.utils.data.DataLoader(test_data_set, batch_size=parameter['batch_size'], shuffle=False,
drop_last=False,
collate_fn=test_data_set.my_collate, num_workers=num_workers)
for ml_epochs in range(parameter['num_epochs']):
lr = learning_rate_scheduler.get_lr(iteration=ml_epochs)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# save loss and labels within an epoch to log the average across the epoch
current_train_loss, current_test_loss = 0, 0
current_train_loss_bert, current_test_loss_bert = 0, 0
true_labels_train = torch.empty(0, dtype=torch.float).to(device)
pred_labels_train = torch.empty(0, dtype=torch.float).to(device)
true_labels_test = torch.empty(0, dtype=torch.float).to(device)
pred_labels_test = torch.empty(0, dtype=torch.float).to(device)
# --------------------------------------------------------------------------------------------------------------
for i, data in enumerate(train_generator):
num_trials = sum([x.size(0) for x in data['patched_eeg_token']])
model.train()
optimizer.zero_grad()
logits = torch.empty(0, n_classes).to(device)
target = torch.empty(0)
current_help_loss = 0
# loop over mini batches with same number of tokens
for sub_batch in range(len(data['patched_eeg_token'])):
transformer_out, logits1, pos_masking = model.forward(x=data['patched_eeg_token'][sub_batch].to(device),
pos=data['pos_as_int'][sub_batch].type(
torch.LongTensor).to(device))
if parameter['pre_train_bert']:
if True in pos_masking:
loss = model.cos_sim_loss(output=
data['patched_eeg_token'][sub_batch][:, 1:].to(device)[pos_masking],
target=transformer_out[pos_masking]) * (
data['labels'][sub_batch].size(0) / num_trials)
loss.backward()
logits = torch.cat((logits, logits1), dim=0)
target = torch.cat((target, data['labels'][sub_batch]), dim=0)
current_help_loss += loss
else:
# scale loss depending of the number of train-trials
loss = criterion(logits1, data['labels'][sub_batch].type(torch.LongTensor).to(device)) * (
data['labels'][sub_batch].size(0) / num_trials)
loss.backward()
logits = torch.cat((logits, logits1), dim=0)
target = torch.cat((target, data['labels'][sub_batch]), dim=0)
current_help_loss += loss
current_train_loss += current_help_loss
'''
if parameter['bert_supervised']:
# add a reconstruction task (BERT -> https://arxiv.org/abs/1810.04805) as regularisation to loss
loss1 = criterion(logits, data['labels'].to(device))
loss2 = model.cos_sim_loss(output=data['brain_epochs'][pos_masking.to(device)],
target=transformer_out[pos_masking.to(device)].to(device))
loss = loss1 + 0.3 * loss2
loss.backward()
current_train_loss += loss1
current_train_loss_bert += loss2
'''
if parameter['clip_gradient']:
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=parameter['clip_gradient'])
optimizer.step()
pred_labels_train = torch.cat((pred_labels_train, logits.argmax(dim=1)), 0)
true_labels_train = torch.cat((true_labels_train, target.to(device)), 0)
# --------------------------------------------------------------------------------------------------------------
# Evaluation step
for j, data in enumerate(test_generator):
num_trials = sum([x.size(0) for x in data['patched_eeg_token']])
with torch.no_grad():
model.eval()
logits = torch.empty(0, n_classes).to(device)
target = torch.empty(0)
current_help_loss = 0
for sub_batch in range(len(data['patched_eeg_token'])):
transformer_out, logits1, pos_masking = model.forward(x=data['patched_eeg_token'][sub_batch]
.to(device),
pos=data['pos_as_int'][sub_batch].type(
torch.LongTensor).to(device))
if parameter['pre_train_bert']:
if True in pos_masking:
loss = model.cos_sim_loss(
output=data['patched_eeg_token'][sub_batch][:, 1:].to(device)[pos_masking],
target=transformer_out[pos_masking]) * (
data['labels'][sub_batch].size(0) / num_trials)
logits = torch.cat((logits, logits1), dim=0)
target = torch.cat((target, data['labels'][sub_batch]), dim=0)
current_help_loss += loss
else:
# scale loss depending of the number of train-trials
loss = criterion(logits1, data['labels'][sub_batch].type(torch.LongTensor).to(device)) * (
data['labels'][sub_batch].size(0) / num_trials)
logits = torch.cat((logits, logits1), dim=0)
target = torch.cat((target, data['labels'][sub_batch]), dim=0)
current_help_loss += loss
current_test_loss += current_help_loss
pred_labels_test = torch.cat((pred_labels_test, logits.argmax(dim=1)), 0)
true_labels_test = torch.cat((true_labels_test, target.to(device)), 0)
current_train_loss = current_train_loss.cpu().detach().numpy() / (i + 1)
current_test_loss = current_test_loss.cpu().detach().numpy() / (j + 1)
if parameter['bert_supervised']:
current_train_loss_bert = current_train_loss_bert.cpu().detach().numpy() / (i + 1)
current_test_loss_bert = current_test_loss_bert.cpu().detach().numpy() / (j + 1)
if not parameter['pre_train_bert']:
current_acc_train = torch.sum(true_labels_train == pred_labels_train) / true_labels_train.size(0)
current_acc_test = torch.sum(true_labels_test == pred_labels_test) / true_labels_test.size(0)
if parameter['wandb_log']:
if parameter['bert_supervised']:
wandb.log({"train_loss": current_train_loss, "test_loss": current_test_loss,
"train_loss_BERT": current_train_loss_bert, "test_loss_BERT": current_test_loss_bert,
'acc_train': current_acc_train, 'acc_test': current_acc_test})
elif parameter['pre_train_bert']:
wandb.log({"train_loss": current_train_loss, "test_loss": current_test_loss,
})
else:
wandb.log({"train_loss": current_train_loss, "test_loss": current_test_loss,
'acc_train': current_acc_train, 'acc_test': current_acc_test,
})
else:
print('')
print('Epoch: {}, train_loss {}, test_loss {}'.format(ml_epochs, current_train_loss, current_test_loss))
if not parameter['pre_train_bert']:
print('Epoch: {}, train_acc {}, test_acc {}'.format(ml_epochs, current_acc_train, current_acc_test))
if parameter['bert_supervised']:
print('Epoch: {}, train_loss_BERT {}, test_loss_BERT {}'.format(ml_epochs, current_acc_train,
current_acc_test))
if parameter['checkpoints']:
if parameter['save']:
if parameter['wandb_log']:
path = os.path.join(parameter['save'], wandb.run.name)
else:
path = parameter['save']
if not os.path.isdir(path):
os.mkdir(path)
if ml_epochs % parameter['checkpoints'] == 0 and ml_epochs != 0:
if not parameter['pre_train_bert']:
torch.save({
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'learning_rate_scheduler': learning_rate_scheduler.state_dict(),
'config': parameter,
'epoch': ml_epochs,
'train_loss': current_train_loss,
'test_loss': current_test_loss,
'acc_train': current_acc_train,
'acc_test': current_acc_test,
'dict_channels': train_data_set.dict_channels,
}, os.path.join(path, 'checkpoint_' + str(ml_epochs) + '.pt'))
else:
torch.save({
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'learning_rate_scheduler': learning_rate_scheduler.state_dict(),
'config': parameter,
'epoch': ml_epochs,
'train_loss': current_train_loss,
'test_loss': current_test_loss,
'dict_channels': train_data_set.dict_channels,
}, os.path.join(path, 'checkpoint_' + str(ml_epochs) + '.pt'))
else:
warnings.warn('Checkpoints are not saved!')
if parameter['save']:
if parameter['wandb_log']:
path = os.path.join(parameter['save'], wandb.run.name)
else:
path = parameter['save']
if not os.path.isdir(path):
os.mkdir(path)
if not parameter['pre_train_bert']:
torch.save({
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'learning_rate_scheduler': learning_rate_scheduler.state_dict(),
'config': parameter,
'epoch': ml_epochs,
'train_loss': current_train_loss,
'test_loss': current_test_loss,
'acc_train': current_acc_train,
'acc_test': current_acc_test,
'dict_channels': train_data_set.dict_channels,
}, os.path.join(path, 'final_' + str(ml_epochs) + '.pt'))
else:
torch.save({
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'learning_rate_scheduler': learning_rate_scheduler.state_dict(),
'config': parameter,
'epoch': ml_epochs,
'train_loss': current_train_loss,
'test_loss': current_test_loss,
'dict_channels': train_data_set.dict_channels,
}, os.path.join(path, 'final_' + str(ml_epochs) + '.pt'))
if parameter['wandb_log']:
wandb.finish()