-
Notifications
You must be signed in to change notification settings - Fork 1
/
main.py
368 lines (310 loc) · 16.6 KB
/
main.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
import numpy as np
import random
from tqdm import tqdm, trange
import os
# Specify CUDA_VISIBLE_DEVICES in the command,
# e.g., CUDA_VISIBLE_DEVICES=0,1 nohup bash exp_on_b7server_0.sh
# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1"
import time
import json
import warnings
warnings.filterwarnings('ignore')
import wandb
import torch
import torch.nn as nn
import torch.nn.functional as F
import transformers
from collections import OrderedDict
from torch.cuda.amp import GradScaler, autocast
from utils.parser_utils import get_args
from utils.logger_utils import get_logger
from utils.other_utils import *
from utils.optimization_utils import *
from utils.mixout_utils import *
from modeling.bert_models import *
def evaluate_accuracy(dev_loader, model):
n_corrects_acm_eval, n_samples_acm_eval = 0.0, 0.0
model.eval()
with torch.no_grad():
num_batch = len(dev_loader)
for batch_idx in tqdm(list(range(num_batch)),total=num_batch,desc='Evaluation'):
input_data = dev_loader[batch_idx]
labels = input_data['example_label']
logits = model.predict(input_data)
bs = logits.shape[0]
n_corrects = n_corrects = (logits.argmax(1) == labels).sum().item()
n_corrects_acm_eval += n_corrects
n_samples_acm_eval += bs
ave_acc_eval = n_corrects_acm_eval / n_samples_acm_eval
return ave_acc_eval
def set_random_seed(seed):
if not seed is None:
logger.info("Fix random seed")
os.environ['PYTHONHASHSEED'] = str(seed)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
# torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
logger.info("Use Random Seed")
def set_wandb(args):
wandb_mode = "online" if args.use_wandb and (not args.debug) else "disabled"
resume = (args.continue_train_from_check_path is not None) and (args.resume_id != "None" and args.resume_id is not None)
args.wandb_id = args.resume_id if resume else wandb.util.generate_id()
args.hf_version = transformers.__version__
wandb_log = wandb.init(mode=wandb_mode, entity="your-entity", project="your-project", config=args, name=args.run_name, resume="allow", id=args.wandb_id, settings=wandb.Settings(start_method="fork"))
logger.info('{0:>30}: {1}'.format("wandb id", args.wandb_id))
return wandb_log
def main(args):
set_random_seed(args.seed)
print_system_info()
print_basic_info(args)
wandb_log = set_wandb(args)
train(args,wandb_log)
def train(args, wandb_log):
logger.info('=' * 71)
logger.info('Start Training')
logger.info('=' * 71)
###################################################################################################
# Get available GPU devices #
###################################################################################################
assert torch.cuda.is_available() and torch.cuda.device_count()>=1, 'No gpu avaliable!'
# Note: Only using the pre-defined gpu_idx when debug; Otherwise, use CUDA_VISIBLE_DEVICES to specify the devices
if (not args.use_wandb) and (args.gpu_idx is not None):
gpu_idx = args.gpu_idx
if isinstance(gpu_idx,int) or (isinstance(gpu_idx,str) and gpu_idx.isdigit()):
devices = torch.device(gpu_idx)
else:
raise Exception('Invalid gpu_idx {gpu_idx}')
else:
# logger.info('{0:>30}: {1}'.format('Visible GPU count',torch.cuda.device_count()))
devices = torch.device(0)
###################################################################################################
# Build model #
###################################################################################################
logger.info("Build model")
if 'bert' in args.pretrain_model:
model = BERT_basic(args)
else:
raise Exception('Invalid pretrain_model name %s'%args.pretrain_model)
# Re-Init
if args.is_ReInit:
# First: Obtain a fully randomly initialized pretrained model
random_init_pretrain_model = deepcopy(model.pretrain_model)
random_init_pretrain_model.apply(random_init_pretrain_model._init_weights) # using apply() to init each submodule recursively
# Then: Set the top layers in the pretrained model
if hasattr(random_init_pretrain_model.config,'num_layers'):
num_layers = random_init_pretrain_model.config.num_layers
elif hasattr(random_init_pretrain_model.config,'num_hidden_layers'):
num_layers = random_init_pretrain_model.config.num_hidden_layers
else:
raise Exception('Cannot find number of layers in model.configs!!!')
ignore_layers = [layer_i for layer_i in range(num_layers-args.ReInit_topk_layer)]
reinit_lst = []
for _name, _para in model.pretrain_model.named_parameters():
# Word embedding don't need initialization
if 'shared' in _name or 'embeddings' in _name:
continue
# for bert
if 'layer.' in _name:
start_idx = _name.find('layer.') +len('layer.')
end_idx = _name.find('.', start_idx)
layer_id = int(_name[start_idx:end_idx])
if layer_id in ignore_layers:
continue
model.pretrain_model.state_dict()[_name][:] = random_init_pretrain_model.state_dict()[_name][:]
reinit_lst.append(_name)
logger.info('Reinit modules: %s'%reinit_lst)
del random_init_pretrain_model
# NoisyTune
if args.is_NoisyTune:
for _name, _para in model.pretrain_model.named_parameters():
model.pretrain_model.state_dict()[_name][:] += (torch.rand(_para.size())-0.5)*args.NoisyTune_lambda*torch.std(_para)
# Mixout
if args.is_Mixout:
# use tuple to avoid OrderedDict warning
for name, module in tuple(model.pretrain_model.named_modules()):
if name:
recursive_setattr(model.pretrain_model, name, replace_layer_for_mixout(module, mixout_prob=args.Mixout_prob))
logger.info('Parameters statistics')
params_statistic(model)
###################################################################################################
# Resume from checkpoint #
###################################################################################################
start_epoch=0
checkpoint_path = os.path.join(args.save_dir, 'checkpoint.pt')
if args.continue_train_from_check_path is not None and args.continue_train_from_check_path != 'None':
logger.info("Resume from checkpoint %s"%args.continue_train_from_check_path)
if torch.cuda.is_available():
check = torch.load(args.continue_train_from_check_path)
else:
check = torch.load(args.continue_train_from_check_path,map_location=torch.device('cpu'))
model_state_dict, _ = check
model.load_state_dict(model_state_dict)
model.train()
###################################################################################################
# Load data #
###################################################################################################
logger.info("Load dataset and dataloader")
dataset = Basic_Dataloader(args, devices=devices)
dev_loader = dataset.dev()
test_loader = dataset.test()
train_loader = dataset.train()
###################################################################################################
# Build Optimizer #
###################################################################################################
logger.info("Build optimizer")
# You can use DataParallel here
# model.pretrain_model = nn.DataParallel(model.pretrain_model, device_ids=(0,1))
# model.pretrain_model.to(devices)
optimizer, scheduler = get_optimizer(model, args, dataset)
# ChildTune
if args.optim == 'childtuningadamw' and args.ChildTuning_mode == 'ChildTuning-D':
model = model.to(devices)
gradient_mask = calculate_fisher(args, model, train_loader)
optimizer.set_gradient_mask(gradient_mask)
model = model.cpu()
###################################################################################################
# Training #
###################################################################################################
model.train()
freeze_net(model.pretrain_model)
logger.info("Freeze model.pretrain_model")
model.to(devices)
# record variables
dev_acc = 0
global_step, best_dev_epoch = 0, 0
best_dev_acc, final_test_acc, best_test_acc = 0.0, 0.0, 0.0
total_loss_acm, n_corrects_acm, n_samples_acm = 0.0, 0.0, 0.0
best_dev_acc = dev_acc
is_finish = False
accumulate_batch_num = args.accumulate_batch_size//args.batch_size
if args.is_CET:
train_loader.generate_refs(model=model, load_cache=True)
for epoch_id in trange(start_epoch, args.n_epochs, desc="Epoch"):
model.epoch_idx = epoch_id
if is_finish:
break
if epoch_id == args.unfreeze_epoch:
unfreeze_net(model.pretrain_model)
logger.info("Unfreeze model.pretrain_model")
if epoch_id == args.refreeze_epoch:
freeze_net(model.pretrain_model)
logger.info("Freeze model.pretrain_model")
model.train()
start_time = time.time()
num_batch = len(train_loader)-1 if args.is_skip_last_batch else len(train_loader)
for batch_id in tqdm(range(num_batch), total=num_batch, desc="Batch"):
# load data for one batch
input_data = train_loader.__getitem__(batch_id, is_skip_last_batch=args.is_skip_last_batch)
labels = input_data['example_label']
bs = len(input_data['example_id'])
if args.is_CET:
loss, logits = model.compute_CET_loss(input_data, labels)
elif args.is_BSS:
loss, logits = model.compute_BSS_loss(input_data, labels)
elif args.is_R3F:
loss, logits = model.compute_R3F_loss(input_data, labels)
else:
loss, logits = model.compute_loss(input_data, labels)
total_loss_acm += loss.item()*bs
loss.requires_grad_(True)
loss.backward()
n_corrects = (logits.detach().argmax(1) == labels).sum().item() if logits is not None else 0
n_corrects_acm += n_corrects
n_samples_acm += bs
if (batch_id+1)%accumulate_batch_num==0 or batch_id==num_batch-1:
if args.max_grad_norm > 0:
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
optimizer.zero_grad()
scheduler.step()
if (global_step + 1) % args.log_interval == 0:
ms_per_batch = 1000 * (time.time() - start_time) / args.log_interval
logger.info('| step {:5} | lr: {:9.7f} | loss {:7.4f} | ms/batch {:7.2f} |'.format(global_step+1, scheduler.get_last_lr()[0], total_loss_acm / n_samples_acm, ms_per_batch))
if not args.debug:
wandb_log.log({"lr": scheduler.get_last_lr()[0], "train_loss": total_loss_acm / n_samples_acm, "train_acc": n_corrects_acm / n_samples_acm, "ms_per_batch": ms_per_batch}, step=global_step+1)
total_loss_acm = 0.0
n_samples_acm = n_corrects_acm = 0
start_time = time.time()
global_step += 1
if epoch_id%args.eval_interval==0:
model.eval()
dev_acc = evaluate_accuracy(dev_loader, model)
test_acc = 0.0
total_acc = []
preds_path = os.path.join(args.save_dir, 'test_e{}_preds.csv'.format(epoch_id))
with open(preds_path, 'w') as f_preds:
with torch.no_grad():
num_batch = len(test_loader)
for batch_idx in tqdm(list(range(num_batch)),total=num_batch,desc='Testing'):
input_data = test_loader[batch_idx]
qids = input_data['example_id']
labels = input_data['example_label']
logits = model.predict(input_data)
predictions = logits.argmax(1) #[bsize, ]
# preds_ranked = (-logits).argsort(1) #[bsize, n_choices]
for i, (qid, label, pred) in enumerate(zip(qids, labels, predictions)):
acc = int(pred.item()==label.item())
f_preds.write('{},{}\n'.format(qid, chr(ord('A') + pred.item())))
f_preds.flush()
total_acc.append(acc)
test_acc = float(sum(total_acc))/len(total_acc)
best_test_acc = max(test_acc, best_test_acc)
if epoch_id >= args.unfreeze_epoch:
# update record variables
if dev_acc >= best_dev_acc:
best_dev_acc = dev_acc
final_test_acc = test_acc
best_dev_epoch = epoch_id
if args.save_model:
model_path = os.path.join(args.save_dir, 'model.pt')
torch.save([model.state_dict(), args], model_path)
logger.info("model saved to %s"%model_path)
else:
best_dev_epoch = epoch_id
logger.info('-' * 71)
logger.info(
'| epoch {:3} | step {:5} | dev_acc {:7.4f} | test_acc {:7.4f} |'.format(epoch_id, global_step, dev_acc,
test_acc))
logger.info('| best_dev_epoch {:3} | best_dev_acc {:7.4f} | final_test_acc {:7.4f} |'.format(best_dev_epoch,
best_dev_acc,
final_test_acc))
logger.info('-' * 71)
if not args.debug:
wandb_log.log({"dev_acc": dev_acc, "dev_loss": dev_acc, "best_dev_acc": best_dev_acc,
"best_dev_epoch": best_dev_epoch}, step=global_step)
if test_acc > 0:
wandb_log.log({"test_acc": test_acc, "test_loss": 0.0, "final_test_acc": final_test_acc},
step=global_step)
if args.save_check:
training_dict = {'epoch':epoch_id, 'loss':loss,
'model_state_dict':model.state_dict(),
'optimizer_state_dict':optimizer.state_dict(),
'scheduler_dict':scheduler.state_dict()}
torch.save(training_dict, checkpoint_path)
if epoch_id - best_dev_epoch >= args.max_epochs_before_stop:
logger.info("After %d epoch no improving. Stop!"%(epoch_id-best_dev_epoch))
logger.info("Best test accuracy: %s"%str(best_test_acc))
logger.info("Final best test accuracy according to dev: %s"%str(final_test_acc))
is_finish=True
break
model.train()
###################################################################################################
# Testing #
###################################################################################################
if args.n_epochs <= 0:
logger.info('n_epochs <= 0, start testing ...')
model.eval()
with torch.no_grad():
dev_acc = evaluate_accuracy(dev_loader, model)
test_acc = evaluate_accuracy(test_loader, model)
logger.info( 'dev_acc {:7.4f} | test_acc {:7.4f}'.format(dev_acc, test_acc))
if __name__ == '__main__':
args = get_args(is_save=True)
logger = get_logger(args)
main(args)