forked from HappyColor/SpeechFormer2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_model.py
357 lines (301 loc) · 16 KB
/
train_model.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Thu Nov 11 18:16:12 CST 2021
@author: lab-chen.weidong
"""
import os
from tqdm import tqdm
import torch
import torch.optim.lr_scheduler as lr_scheduler
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.utils.tensorboard import SummaryWriter
from collections import OrderedDict
import json
import argparse
import utils
from config import Cfg, create_workshop, modify_config
class Engine():
def __init__(self, cfg, local_rank: int, world_size: int):
self.cfg = cfg
self.local_rank = local_rank
self.world_size = world_size
self.ckpt_save_path = self.cfg.ckpt_save_path
self.device = self.cfg.train.device
self.EPOCH = self.cfg.train.EPOCH
self.current_epoch = 0
self.iteration = 0
if self.cfg.train.find_init_lr:
self.cfg.train.lr = 0.000001
self.cfg.train.step_size = 1
self.cfg.train.gamma = 1.05
if self.local_rank == 0:
self.writer = SummaryWriter(self.cfg.workshop)
### prepare model and train tools
model_type = self.cfg.model.type
model_type = 'SpeechFormer_v2' if model_type == 'SpeechFormer++' else model_type
with open('./config/model_config.json', 'r') as f1, open(f'./config/{self.cfg.dataset.database}_feature_config.json', 'r') as f2:
model_json = json.load(f1)[model_type]
feas_json = json.load(f2)
data_json = feas_json[self.cfg.dataset.feature]
data_json['meta_csv_file'] = feas_json['meta_csv_file']
model_json['num_classes'] = feas_json['num_classes']
model_json['input_dim'] = (data_json['feature_dim'] // model_json['num_heads']) * model_json['num_heads']
model_json['length'] = data_json['length']
model_json['ffn_embed_dim'] = model_json['input_dim'] // 2
model_json['hop'] = data_json['hop']
self.model = utils.model.load_model(model_type, self.device, **model_json)
self.optimizer = torch.optim.SGD(filter(lambda x: x.requires_grad, self.model.parameters()), lr=self.cfg.train.lr, momentum=0.9)
self.loss_func = torch.nn.CrossEntropyLoss()
self.calculate_score = utils.toolbox.calculate_score_classification
# self.data_loader_feactory = utils.dataset.DataloaderFactory(self.cfg)
self.data_loader_feactory = utils.dataset_lmdb.DataloaderFactory(self.cfg)
self.train_dataloader = self.data_loader_feactory.build(state='train', **data_json)
self.test_dataloader = self.data_loader_feactory.build(state='test', **data_json)
if self.cfg.train.find_init_lr:
self.scheduler = lr_scheduler.StepLR(self.optimizer, step_size=self.cfg.train.step_size, gamma=self.cfg.train.gamma)
else:
self.scheduler = lr_scheduler.CosineAnnealingLR(optimizer=self.optimizer, T_max=self.EPOCH, eta_min=self.cfg.train.lr / 100)
### prepare logger
self.logger_train = utils.logger.create_logger(self.cfg.workshop, name='train') if self.local_rank == 0 else None
self.logger_test = utils.logger.create_logger(self.cfg.workshop, name='test') if self.local_rank == 0 else None
if self.logger_train is not None:
self.logger_train.info(f'workshop: {self.cfg.workshop}')
self.logger_train.info(f'seed: {self.cfg.train.seed}')
self.logger_train.info(f'pid: {os.getpid()}')
### prepare meters
data_type = torch.int64
self.loss_meter = utils.avgmeter.AverageMeter(device='cpu')
self.score_meter = utils.avgmeter.AverageMeter(device='cpu')
self.predict_recoder = utils.recoder.TensorRecorder(device='cuda', dtype=data_type)
self.label_recoder = utils.recoder.TensorRecorder(device='cuda', dtype=data_type)
self.tag_recoder = utils.recoder.StrRecorder(device='cpu', dtype=str)
self.train_score_1, self.train_score_2, self.train_score_3, self.train_loss = [], [], [], []
self.test_score_1, self.test_score_2, self.test_score_3, self.test_loss = [], [], [], []
def reset_meters(self):
self.loss_meter.reset()
self.score_meter.reset()
def reset_recoders(self):
self.predict_recoder.reset()
self.label_recoder.reset()
self.tag_recoder.reset()
def gather_distributed_data(self, gather_data):
if isinstance(gather_data, torch.Tensor):
_output = [torch.zeros_like(gather_data) for _ in range(self.world_size)]
dist.all_gather(_output, gather_data, async_op=False)
output = torch.cat(_output)
else:
if gather_data[0] is not None:
_output = [None for _ in range(self.world_size)]
if hasattr(dist, 'all_gather_object'):
dist.all_gather_object(_output, gather_data)
else:
utils.distributed.all_gather_object(_output, gather_data, self.world_size)
output = []
for lst in _output:
output.extend(lst)
else:
output = None
return output
def train_epoch(self):
self.train_dataloader.set_epoch(self.current_epoch)
if self.local_rank == 0:
print(f'-------- {self.cfg.workshop} --------')
discrip_str = f'Epoch-{self.current_epoch}/{self.EPOCH}'
pbar_train = tqdm(self.train_dataloader, disable=self.local_rank != 0)
pbar_train.set_description('Train' + discrip_str)
self.reset_meters()
self.reset_recoders()
self.model.train()
for data in pbar_train:
self.iteration += 1
x = torch.stack(data[0], dim=0).to(self.device)
y = torch.tensor(data[1]).to(self.device)
vote_tag = data[2]
batch_size = y.shape[0]
self.optimizer.zero_grad()
out = self.model(x)
loss = self.loss_func(out, y)
loss.backward()
self.optimizer.step()
y_pred = torch.argmax(out, dim=1)
self.predict_recoder.record(y_pred)
self.label_recoder.record(y)
self.tag_recoder.record(vote_tag)
score = utils.toolbox.calculate_basic_score(y_pred.cpu(), y.cpu())
self.loss_meter.update(loss.item())
self.score_meter.update(score, batch_size)
pbar_train_dic = OrderedDict()
pbar_train_dic['iter'] = self.iteration
pbar_train_dic['lr'] = self.optimizer.param_groups[0]['lr']
pbar_train_dic['score'] = f'{self.score_meter.avg:.5f}' # acc / MSE in local_rank: 0
pbar_train_dic['loss'] = f'{self.loss_meter.avg:.5f}' # loss in local_rank: 0
pbar_train.set_postfix(pbar_train_dic)
if self.cfg.train.find_init_lr:
if loss.item() > 20:
raise ValueError(f'Loss: {loss.item()} started to expand. Please use tensorboard to find the appropriate lr.')
if self.local_rank == 0:
self.writer.add_scalar('Step Loss', loss.item(), self.iteration)
self.writer.add_scalar('Total Loss', self.loss_meter.avg, self.iteration)
self.writer.add_scalar('Step LR', self.optimizer.param_groups[0]['lr'], self.iteration)
self.scheduler.step()
epoch_preds = self.gather_distributed_data(self.predict_recoder.data).cpu()
epoch_labels = self.gather_distributed_data(self.label_recoder.data).cpu()
epoch_tag = self.gather_distributed_data(self.tag_recoder.data)
if self.local_rank == 0:
average_f1 = 'weighted' if self.cfg.dataset.database == 'meld' else 'macro'
score_1, score_2, score_3, score_4, score_5 = self.calculate_score(epoch_preds, epoch_labels, average_f1)
self.train_score_1.append(score_1)
self.train_score_2.append(score_2)
self.train_score_3.append(score_3)
self.train_loss.append(self.loss_meter.avg)
if self.logger_train is not None:
self.logger_train.info(
f'Training epoch: {self.current_epoch}, accuracy: {score_1:.5f}, precision: {score_4:.5f}, recall: {score_2:.5f}, F1: {score_3:.5f}, loss: {self.loss_meter.avg:.5f}'
)
def test(self):
discrip_str = f'Epoch-{self.current_epoch}'
pbar_test = tqdm(self.test_dataloader, disable=self.local_rank != 0)
pbar_test.set_description('Test' + discrip_str)
self.reset_meters()
self.reset_recoders()
self.model.eval()
with torch.no_grad():
for data in pbar_test:
x = torch.stack(data[0], dim=0).to(self.device)
y = torch.tensor(data[1]).to(self.device)
vote_tag = data[2]
batch_size = y.shape[0]
out = self.model(x)
loss = self.loss_func(out, y)
y_pred = torch.argmax(out, dim=1)
self.predict_recoder.record(y_pred)
self.label_recoder.record(y)
self.tag_recoder.record(vote_tag)
score = utils.toolbox.calculate_basic_score(y_pred.cpu(), y.cpu())
self.loss_meter.update(loss.item())
self.score_meter.update(score, batch_size)
pbar_test_dic = OrderedDict()
pbar_test_dic['score'] = f'{self.score_meter.avg:.5f}'
pbar_test_dic['loss'] = f'{self.loss_meter.avg:.5f}'
pbar_test.set_postfix(pbar_test_dic)
epoch_preds = self.gather_distributed_data(self.predict_recoder.data).cpu()
epoch_labels = self.gather_distributed_data(self.label_recoder.data).cpu()
epoch_tag = self.gather_distributed_data(self.tag_recoder.data)
if self.local_rank == 0:
if hasattr(self.cfg.train, 'vote'):
if self.cfg.dataset.database == 'pitt':
modify_tag_func = utils.toolbox._majority_target_Pitt
elif self.cfg.dataset.database == 'daic_woz':
modify_tag_func = utils.toolbox._majority_target_DAIC_WOZ
else:
raise KeyError(f'Database: {self.cfg.dataset.database} do not need voting!')
_, epoch_preds, epoch_labels = utils.toolbox.majority_vote(epoch_tag, epoch_preds, epoch_labels, modify_tag_func)
average_f1 = 'weighted' if self.cfg.dataset.database == 'meld' else 'macro'
# Calculate accuracy, recall, F1, precision, confuse_matrix
score_1, score_2, score_3, score_4, score_5 = self.calculate_score(epoch_preds, epoch_labels, average_f1)
self.test_score_1.append(score_1) # accuracy
self.test_score_2.append(score_2) # recall
self.test_score_3.append(score_3) # F1
self.test_loss.append(self.loss_meter.avg)
if self.cfg.train.save_best:
if self.cfg.dataset.database in ['iemocap', 'pitt']:
is_best = max(self.test_score_1) == score_1 # accuracy is more important
elif self.cfg.dataset.database in ['meld', 'daic_woz']:
is_best = max(self.test_score_3) == score_3 # F1 is more important
else:
is_best = False
if is_best:
self.model_save(is_best=True)
if self.logger_test is not None:
self.logger_test.info(
f'Testing epoch: {self.current_epoch}, accuracy: {score_1:.5f}, precision: {score_4:.5f}, recall: {score_2:.5f}, F1: {score_3:.5f}, loss: {self.loss_meter.avg:.5f}, confuse_matrix: \n{score_5}'
)
def model_save(self, is_best=False):
if is_best:
ckpt_save_file = os.path.join(self.ckpt_save_path, 'best.pt')
else:
ckpt_save_file = os.path.join(self.ckpt_save_path, f'epoch{self.current_epoch}.pt')
save_dict = {
'cfg': self.cfg,
'epoch': self.current_epoch,
'model': self.model.state_dict(),
'optimizer': self.optimizer.state_dict()
}
torch.save(save_dict, ckpt_save_file)
def run(self):
if self.cfg.train.find_init_lr:
while self.current_epoch < self.EPOCH:
self.train_epoch()
self.current_epoch += 1
else:
plot_sub_titles = ['WA-train', 'UA-train', 'F1-train', 'Loss-train', 'WA-test', 'UA-test', 'F1-test', 'Loss-test']
plot_data_name = ['train_score_1', 'train_score_2', 'train_score_3', 'train_loss', 'test_score_1', 'test_score_2', 'test_score_3', 'test_loss']
while self.current_epoch < self.EPOCH:
self.train_epoch()
self.scheduler.step()
self.test()
self.current_epoch += 1
if self.local_rank == 0:
plot_data = [getattr(self, data_name) for data_name in plot_data_name]
utils.write_result.plot_process(plot_data, plot_sub_titles, self.cfg.workshop)
utils.logger.close_logger(self.logger_train)
utils.logger.close_logger(self.logger_test)
def main_worker(local_rank, cfg, world_size, dist_url):
utils.environment.set_seed(cfg.train.seed + local_rank)
torch.cuda.set_device(local_rank)
dist.init_process_group(
backend='nccl',
init_method=dist_url,
world_size=world_size,
rank=local_rank,
)
if cfg.dataset.database == 'iemocap':
cfg.train.strategy = '5cv'
folds = [1, 2, 3, 4, 5] if cfg.train.folds is None else cfg.train.folds
folds = [folds] if not isinstance(folds, list) else folds
elif cfg.dataset.database == 'meld':
folds = [1]
elif cfg.dataset.database == 'pitt':
cfg.train.vote = True
folds = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] if cfg.train.folds is None else cfg.train.folds
folds = [folds] if not isinstance(folds, list) else folds
elif cfg.dataset.database == 'daic_woz':
cfg.train.vote = True
folds = [1]
else:
raise KeyError(f'Unknown database: {cfg.dataset.database}')
for f in folds:
cfg_clone = cfg.clone()
cfg_clone.train.current_fold = f
create_workshop(cfg_clone, local_rank)
engine = Engine(cfg_clone, local_rank, world_size)
engine.run()
torch.cuda.empty_cache()
if local_rank == 0:
criterion = ['accuracy', 'precision', 'recall', 'F1']
evaluate = cfg.train.evaluate
outfile = f'result/result_{cfg.model.type}.csv'
utils.write_result.path_to_csv(os.path.dirname(cfg_clone.workshop), criterion, evaluate, csvfile=outfile)
def main(cfg):
utils.environment.visible_gpus(cfg.train.device_id)
utils.environment.set_seed(cfg.train.seed)
free_port = utils.distributed.find_free_port()
dist_url = f'tcp://127.0.0.1:{free_port}'
world_size = torch.cuda.device_count() # num_gpus
print(f'world_size={world_size} Using dist_url={dist_url}')
mp.spawn(fn=main_worker, args=(cfg, world_size, dist_url), nprocs=world_size)
if __name__=='__main__':
parser = argparse.ArgumentParser()
parser.add_argument("-mo", "--model.type", help="modify cfg.train.model.type", type=str)
parser.add_argument("-d", "--dataset.database", help="modify cfg.dataset.database", type=str)
parser.add_argument("-f", "--dataset.feature", help="modify cfg.dataset.feature", type=str)
parser.add_argument("-g", "--train.device_id", help="modify cfg.train.device_id", type=str)
parser.add_argument("-m", "--mark", help="modify cfg.mark", type=str)
parser.add_argument("-s", "--train.seed", help="modify cfg.train.seed", type=int)
parser.add_argument("-save", "--train.save_best", help="modify cfg.train.save_best", action='store_true')
parser.add_argument("-folds", "--train.folds", nargs='*', help="modify cfg.train.folds", type=int)
args = parser.parse_args()
modify_config(Cfg, args)
main(Cfg)