-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
460 lines (375 loc) · 21.1 KB
/
trainer.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
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
from numpy.lib.function_base import average
from sklearn.utils import shuffle
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils import data
from torch.utils.data import DataLoader
from sklearn.metrics import f1_score, accuracy_score, confusion_matrix
import numpy as np
import pandas as pd
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
import os
from utils import *
from loss.ND_Crossentropy import CrossentropyND, TopKLoss
from loss.focalloss import *
from loss.label_smoothing import LabelSmoothingCrossEntropy
from utils import gpu_checking
from collections import defaultdict
from modAL.utils.selection import multi_argmax, shuffled_argmax
import wandb
class TrainMaker:
def __init__(self, args, model, data, data_v=None):
self.args = args
self.model = model
self.data = data
if data_v:
self.data_v = data_v
self.history = defaultdict(list)
self.writer = SummaryWriter(log_dir='./runs/lr{}_wd{}'.format(self.args.lr, self.args.wd))
self.optimizer = getattr(torch.optim, self.args.optimizer)
self.optimizer = self.optimizer(self.model.parameters(), lr=self.args.lr, weight_decay=self.args.wd)
# self.optimizer = self.args.optimizer
# self.scheduler = optim.lr_scheduler.StepLR(self.optimizer, step_size=500, gamma=0.1)
self.scheduler = self.__set_scheduler(args, self.optimizer)
# criterion = getattr(torch.nn, self.criterion)
self.criterion = self.__set_criterion(self.args.criterion)
self.epoch = self.args.epoch
self.lr = self.args.lr
self.wd = self.args.wd
self.channel_num = self.args.channel_num
self.device = gpu_checking()
self.history_mini_batch = defaultdict(list)
# if args.mode == "train":
# self.trainer = self.__make_trainer(args=args,
# model=self.model,
# data=data,
# criterion=self.criterion,
# optimizer=self.optimizer)
def training(self, shuffle=True, interval=1000):
prev_v = -10
sampler = torch.utils.data.WeightedRandomSampler(self.data.in_weights, replacement=True, num_samples=self.args.batch_size)
# data_loader = train_data
data_loader = DataLoader(self.data, batch_size=self.args.batch_size, sampler=sampler)
# data_loader = DataLoader(self.data, batch_size=self.args.batch_size)
# data_loader = self.data.data_loader_maker(self.args, shuffle, mode="train")
for e in tqdm(range(self.epoch)):
epoch_loss = 0
self.model.train()
pred_label_acc = None
true_label_acc = None
pred_list = None
true_label = None
# history_mini_batch = defaultdict(list)
for idx, data in enumerate(data_loader):
x, y = data
true_label = y.numpy()
b = x.shape[0]
self.optimizer.zero_grad() # optimizer 항상 초기화
#x = x.reshape(b,1,30,-1) ############## checking [1, 1, 30, 700] 미츼ㅣㅣ
x = x.reshape(b, 1, self.channel_num, -1) # [1, 1, 25, 750]
pred = self.model(x.to(device=self.device).float())
pred_prob = F.softmax(pred, dim=-1)
# pred, (hidden_state, cell_state) = network(x.view(b, 1, -1).float().to(device=device), (hidden_state[:,:b].detach().contiguous(), cell_state[:,:b].detach().contiguous())) # view함수 -> 밑에 적음
loss = self.criterion(pred_prob, y.flatten().long().to(device=self.device))
if (idx+1) % interval == 0: print('[Epoch{}, Step({}/{})] Loss:{:.4f}'.format(e+1, idx+1, len(data_loader.dataset) // self.args.batch_size, epoch_loss / (idx + 1)))
loss.backward()
epoch_loss += loss.item()
self.optimizer.step()
pred_label = torch.argmax(pred_prob, dim=-1).cpu().numpy()
# pred_list.append(pred_label) # pred_list에 prediction 넣어주기
if pred_label_acc is None:
pred_label_acc = pred_label
true_label_acc = true_label
else:
pred_label_acc = np.concatenate((pred_label_acc, pred_label), axis=None)
true_label_acc = np.concatenate((true_label_acc, true_label), axis=None)
# Calculate log per mini-batch
# log = calculate(self.args.metrics, loss, labels, outputs, acc_count=True)
self.cal = Calculate()
log = self.cal.calculator(metrics=self.args.metrics, loss=loss, y_true=y, y_pred=pred_label, acc_count=True)
# Record history per mini-batch
self.record_history(log, self.history_mini_batch, phase='train')
# pred_list_acc = np.concatenate((pred_list_acc, pred_list))
# print(pred_list_acc)
# print("=====", true_label)
# print("true_label:", true_label)
# print(true_label)
# print("\n\n\n", pred_list)
# print(true_label_acc)
# print(pred_label_acc)
f1 = f1_score(true_label_acc, pred_label_acc, average='macro')
acc = accuracy_score(true_label_acc, pred_label_acc)
cm = confusion_matrix(true_label_acc, pred_label_acc)
epoch_loss = epoch_loss / (idx+1)
# print('\nEpoch{} Training, f1:{:.4f}, acc:{:.4f}, Loss:{:.4f}'.format(e+1, f1, acc, epoch_loss))
# print(cm) # confusion matrix print
f1_v, acc_v, cm_v, loss_v = self.evaluation(self.data_v)
if f1_v > prev_v :
prev_v = f1_v
create_folder('./param/lr{}_wd{}'.format(self.lr, self.wd))
torch.save(self.model.state_dict(), './param/lr{}_wd{}/eegnet_f1_{:.2f}'.format(self.lr, self.wd, f1_v))
self.save_checkpoint(epoch=len(self.history['train_loss']))
self.write_history(self.history_mini_batch)
# writer.add_scalar('Learning Rate', lr.get_last_lr()[-1], e)
self.writer.add_scalar('Train/Loss', epoch_loss, e)
self.writer.add_scalar('Train/F1', f1, e)
self.writer.add_scalar('Train/Acc', acc, e)
self.writer.add_scalar('Valid/Loss', loss_v, e)
self.writer.add_scalar('Valid/F1', f1_v, e)
self.writer.add_scalar('Valid/Acc', acc_v, e)
self.writer.flush()
# wandb.log({"loss": epoch_loss,
# "acc": acc,
# "f1":f1,
# "vloss": loss_v,
# "vacc": acc_v,
# "vf1":f1_v,
# "lr": self.optimizer.state_dict().get('param_groups')[0].get("lr")
# })
self.scheduler.step()
return acc, f1, cm, epoch_loss
def evaluation(self, data, interval=1000):
data_loader = DataLoader(data, batch_size=self.args.batch_size)
with torch.no_grad(): # gradient 안함
self.model.eval() # dropout은 training일 때만, evaluation으로 하면 dropout 해제
pred_label_acc = None
true_label_acc = None
pred_list = None
true_label = None
valid_loss = 0
for idx, data in enumerate(data_loader):
x, y = data
true_label = y.numpy()
b = x.shape[0]
x = x.reshape(b, 1, self.channel_num, -1)
pred = self.model(x.to(device=self.device).float())
# 밑에 두개 중에뭐지?
# pred = self.model(x.transpose(1,2).reshape(b,1,self.channel_num,-1).to(device=self.device).float())
#pred = self.model(x.reshape(b,1,30,-1).to(device=device).float()) #self.model(x.transpose(1,2).reshape(b,1,30,-1).to(device=device).float())
loss = self.criterion(pred, y.flatten().long().to(device=self.device)) # pred.shape
valid_loss += loss
# if (idx+1) % interval == 0: print('[Epoch, Step({}/{})] Valid Loss:{:.4f}'.format(idx+1, len(data)//self.args.batch_size, loss / (idx +1)))
pred_prob = F.softmax(pred, dim=-1)
pred_label = torch.argmax(pred_prob, dim = -1).cpu().numpy()
# print(pred_prob)
if pred_label_acc is None:
pred_label_acc = pred_label
true_label_acc = true_label
else:
pred_label_acc = np.concatenate((pred_label_acc, pred_label), axis=None)
true_label_acc = np.concatenate((true_label_acc, true_label), axis=None)
self.cal = Calculate()
log = self.cal.calculator(metrics=self.args.metrics, loss=loss, y_true=y, y_pred=pred_label, acc_count=True)
# Record history per mini-batch
self.record_history(log, self.history_mini_batch, phase='val')
# valid_loss = valid_loss / (idx+1)
# pred_list = np.concatenate(pred_list)
# true_label = np.concatenate(true_label)
# print(true_label_acc)
# print(pred_label_acc)
f1 = f1_score(true_label_acc, pred_label_acc, average='macro')
acc = accuracy_score(true_label_acc, pred_label_acc)
cm = confusion_matrix(true_label_acc, pred_label_acc)
if not self.args.mode == "test":
print('\nEpoch Validation, f1:{:.4f}, acc:{:.4f}, Loss:{:.4f}'.format(f1, acc, valid_loss))
else:
print('\nEpoch Test, f1:{:.4f}, acc:{:.4f}'.format(f1, acc))
# print(cm)
return f1, acc, cm, valid_loss
def predict_proba(self, data, interval=1000, n_instances=1, mcdo=False, random=False, *query_args, **query_kwargs):
data_loader = DataLoader(data, batch_size=self.args.batch_size)
if not mcdo:
with torch.no_grad(): # gradient 안함
self.model.eval() # dropout은 training일 때만, evaluation으로 하면 dropout 해제 ############################
pred_list = []
uncertainty_list = []
for idx, data in enumerate(data_loader):
# x = data
# true_label.append(y)
x = data
b = x.shape[0]
x = x.reshape(b, 1, self.channel_num, -1)
pred = self.model(x.to(device=self.device).float())
# pred = self.model(x.transpose(1,2).reshape(b,1,self.channel_num,-1).to(device=self.device).float())
pred_prob = F.softmax(pred, dim=-1)
# uncertainty = torch.max(pred_prob, dim=-1).values.cpu()
uncertainty = torch.ones(pred_prob.shape[0]).cpu() - torch.max(pred_prob, dim=-1).values.cpu()
pred_label = torch.argmax(pred_prob, dim = -1).cpu().numpy()
pred_list.append(pred_label)
uncertainty_list.append(uncertainty)
pred_list = np.concatenate(pred_list)
uncertainty_list = np.concatenate(uncertainty_list)
if random is True:
print("[Random strategy]")
max_idx = np.random.choice(data.shape[0], 1, False)
else:
max_idx = multi_argmax(uncertainty_list, n_instances=n_instances)
# print(pred_list)
pseudo_labeling = pred_list[max_idx]
# return multi_argmax(uncertainty_list, n_instances=query_kwargs['n_instances'])
else:
with torch.no_grad():
self.model.eval()
self.enable_dropout(self.model.clf)
uncertainty_list = []
mean_lists = []
for idx, data in enumerate(data_loader):
pred_list = []
for _ in range(10) :
# x = data
# true_label.append(y)
x = data
b = x.shape[0]
x = x.reshape(b, 1, self.channel_num, -1)
pred = self.model(x.to(device=self.device).float())
# pred = self.model(x.transpose(1,2).reshape(b,1,self.channel_num,-1).to(device=self.device).float())
pred_prob = F.softmax(pred, dim=-1)
# uncertainty = torch.max(pred_prob, dim=-1).values.cpu()
pred_label = torch.argmax(pred_prob, dim = -1).cpu().numpy()
pred_list.append(pred_label)
std_list = np.std(pred_list, axis=0)
mean_list = np.mean(pred_list, axis=0)
# pred_list = np.concatenate(pred_list)
uncertainty_list.append(std_list)
mean_lists.append(mean_list)
uncertainty_list = np.concatenate(uncertainty_list)
pred_list = np.concatenate(mean_lists)
pred_list = pred_list.round(0)
max_idx = multi_argmax(uncertainty_list, n_instances=n_instances)
# print(pred_list)
pseudo_labeling = pred_list[max_idx]
return max_idx, pseudo_labeling
def predict_score(self, data, interval=1000, *query_args, **query_kwargs):
data_loader = DataLoader(data, batch_size=self.args.batch_size)
with torch.no_grad(): # gradient 안함
self.model.eval() # dropout은 training일 때만, evaluation으로 하면 dropout 해제 ############################
pred_list = []
true_label = []
for idx, data in enumerate(data_loader):
x, y = data
true_label.append(y)
b = x.shape[0]
pred = self.model(x.transpose(1,2).reshape(b,1,self.channel_num,-1).to(device=self.device).float())
pred_prob = F.softmax(pred, dim=-1)
pred_label = torch.argmax(pred_prob, dim = -1).cpu().numpy()
pred_list.append(pred_label)
pred_list = np.concatenate(pred_list)
true_label = np.concatenate(true_label)
f1 = f1_score(true_label, pred_list, average="macro")
return f1
def pseudo_label(self, data, interval=1000, *query_args, **query_kwargs):
print("[trainer]pseudo_label에서의 print", data.x.shape)
data_loader = DataLoader(data, batch_size=self.args.batch_size)
pred_list = []
with torch.no_grad(): # gradient 안함
self.model.eval() # dropout은 training일 때만, evaluation으로 하면 dropout 해제 ############################
for idx, data in enumerate(data_loader):
x, y = data
b = x.shape[0]
x = x.reshape(b, 1, self.channel_num, -1)
pred = self.model(x.to(device=self.device).float())
# pred = self.model(x.transpose(1,2).reshape(b,1,self.channel_num,-1).to(device=self.device).float())
pred_prob = F.softmax(pred, dim=-1)
pred_label = torch.argmax(pred_prob, dim = -1).cpu().numpy()
pred_list.append(pred_label)
pred_list = np.concatenate(pred_list)
return pred_list
def record_history(self, log, history, phase):
for metric in log:
history[f'{phase}_{metric}'].append(log[metric])
def write_history(self, history):
for metric in history:
if metric.endswith('acc'):
n_samples = self.data.x.shape[0]
# n_samples = len(getattr(self.data, f"{metric.split('_')[0]}_loader").dataset.y)
self.history[metric].append((sum(history[metric]) / n_samples))
else:
self.history[metric].append(sum(history[metric]) / len(history[metric]))
# if self.args.mode == 'train':
# write_json(os.path.join(self.args.save_path, "history.json"), self.history)
# else:
# write_json(os.path.join(self.args.save_path, "history_test.json"), self.history)
def save_checkpoint(self, epoch):
create_folder(os.path.join(self.args.save_path, "checkpoints"))
torch.save({
'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
# 'scheduler_state_dict': self.scheduler.state_dict() if self.scheduler else None
}, os.path.join(self.args.save_path, f"checkpoints/{epoch}.tar"))
# if epoch >= 6:
# os.remove(os.path.join(self.args.save_path, f"checkpoints/{epoch - 5}.tar"))
def __set_criterion(self, criterion):
if criterion == "MSE":
criterion = nn.MSELoss()
elif criterion == "CEE":
criterion = nn.CrossEntropyLoss()
elif criterion == "Focal":
criterion = FocalLoss(gamma=2)
elif criterion == "ND":
criterion = CrossentropyND()
elif criterion == "TopK":
criterion = TopKLoss()
elif criterion == "LS":
criterion = LabelSmoothingCrossEntropy()
return criterion
def __set_scheduler(self, args, optimizer):
if args.scheduler is None:
return None
elif args.scheduler == 'exp':
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=args.gamma)
elif args.scheduler == 'step':
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=args.step_size, gamma=args.gamma)
elif args.scheduler == 'multi_step':
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.milestones, gamma=args.gamma)
elif args.scheduler == 'plateau':
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=20,
threshold=0.1, threshold_mode='abs', verbose=True)
elif args.scheduler == 'cosine':
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
T_max=args.T_max if args.T_max else args.epochs,
eta_min=args.eta_min if args.eta_min else 0)
else:
raise ValueError(f"Not supported {args.scheduler}.")
return scheduler
def enable_dropout(self, model):
""" Function to enable the dropout layers during test-time """
for m in model.modules():
if m.__class__.__name__.startswith('Dropout'):
m.train()
# def __make_trainer(self, **kwself.args):
# module = importlib.import_module(f"trainers.{kwself.args['self.args'].model}_trainer")
# trainer = getattr(module, 'Trainer')(**kwargs)
# return trainer
class Calculate:
def calculator(self, metrics, loss, y_true, y_pred, numpy=True, **kwargs):
if numpy:
y_true = self.guarantee_numpy(y_true)
y_pred = self.guarantee_numpy(y_pred)
history = defaultdict(list)
for metric in metrics:
history[metric] = getattr(self, f"get_{metric}")(loss=loss, y_true=y_true, y_pred=y_pred, **kwargs)
return history
def get_loss(self, loss, **kwargs):
return float(loss)
def get_acc(self, y_true, y_pred, acc_count=False, **kwargs):
if acc_count:
return sum(y_true == y_pred)
else:
return sum(y_true == y_pred) / len(y_true)
def guarantee_numpy(self, data):
data_type = type(data)
if data_type == torch.Tensor:
device = data.device.type
if device == 'cpu':
data = data.detach().numpy()
else:
data = data.detach().cpu().numpy()
return data
elif data_type == np.ndarray or data_type == list:
return data
else:
raise ValueError("Check your data type.")