-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
213 lines (176 loc) · 9.27 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
import torch
import numpy as np
import random
import os
from configs.options import get_parser
import logging
from src.utils import load_config, init_logging
from src.data.datasets import SignRegTranDataset
from src.models.signmodel_rel import SignModel
from src.data.vocabulary import GlsTextVocab
from src.criterion.ctc_and_ce_loss import CtcCeLoss
from src.trainer import Trainer
from tqdm import tqdm
from metrics.metrics import wer_single
from phoenix_utils.phoenix_cleanup import clean_phoenix_2014_trans, clean_phoenix_2014
from metrics.metrics import bleu, chrf, rouge
import torch.optim as optim
from src.utils import ModelManager, ModelManager_bleu
import time
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
def main():
opts = get_parser()
setup_seed(opts.seed)
if not os.path.exists(opts.log_dir):
os.makedirs(opts.log_dir)
log_name = "_".join(time.asctime(time.localtime(time.time())).split(" ")[-2].split(":"))
if opts.mode == "train" or opts.mode == "dev":
init_logging(os.path.join(opts.log_dir, "train_{}.log".format(log_name)))
else:
init_logging(os.path.join(opts.log_dir, "test_{}.log".format(log_name)))
if torch.cuda.is_available():
# torch.cuda.set_device(opts.gpu)
logging.info("Using GPU!")
device = "cuda"
else:
logging.info("Using CPU!")
device = "cpu"
logging.info(opts)
gls_text_vocab = GlsTextVocab(opts)
if opts.mode == "train":
train_datasets = SignRegTranDataset(opts, gls_text_vocab, phrase="train", DEBUG=opts.DEBUG)
valid_datasets = SignRegTranDataset(opts, gls_text_vocab, phrase="dev", DEBUG=opts.DEBUG, sample=False)
test_datasets = SignRegTranDataset(opts, gls_text_vocab, phrase="test", DEBUG=opts.DEBUG, sample=False)
else:
train_datasets = SignRegTranDataset(opts, gls_text_vocab, phrase="dev", DEBUG=opts.DEBUG, sample=False)
valid_datasets = SignRegTranDataset(opts, gls_text_vocab, phrase="dev", DEBUG=opts.DEBUG, sample=False)
test_datasets = SignRegTranDataset(opts, gls_text_vocab, phrase="test", DEBUG=opts.DEBUG, sample=False)
model = SignModel(opts, gls_text_vocab, do_recognition=True, do_translation=True)
criterion = CtcCeLoss(opts, gls_text_vocab, reg_loss_weight=opts.reg_loss_weight,
tran_loss_weight=opts.tran_loss_weight, smoothing=opts.label_smoothing)
logging.info(model)
# exit()
trainer = Trainer(opts, model, criterion, gls_text_vocab)
# lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(
# trainer.optimizer, factor=opts.decrease_factor, patience=opts.patience)
if os.path.exists(opts.check_point):
logging.info("Loading checkpoint file from {}".format(opts.check_point))
epoch, num_updates, loss = trainer.load_checkpoint(opts.check_point)
else:
logging.info("No checkpoint file in found in {}".format(opts.check_point))
epoch, num_updates, loss = 0, 0, 0.0
logging.info('| num. module params: {} (num. trained: {})'.format(
sum(p.numel() for p in model.parameters()),
sum(p.numel() for p in model.parameters() if p.requires_grad),
))
trainer.set_num_updates(num_updates)
model_manager = ModelManager(max_num_models=5)
model_manager_bleu = ModelManager_bleu(max_num_models=5)
best_wer = 100.0
best_bleu = 0.0
while epoch < opts.max_epoch and trainer.get_num_updates() < opts.max_updates:
epoch += 1
# trainer.slr_lr_schedule(epoch)
if opts.mode == "train" or opts.mode == "dev":
loss = train(opts, train_datasets, trainer, epoch)
reg_res, tran_res = eval(opts, valid_datasets, trainer, epoch, gls_text_vocab)
eval(opts, test_datasets, trainer, epoch, gls_text_vocab)
if reg_res["wer"] < best_wer:
best_wer = reg_res["wer"]
if tran_res["bleu"]["bleu4"] > best_bleu:
best_bleu = tran_res["bleu"]["bleu4"]
logging.info("Epoch: {}, best_wer: {:.4f} and best_bleu-4: {:.4f}".format(
epoch, best_wer, best_bleu
))
save_ckpt = os.path.join(opts.log_dir, 'ep{:d}_wer_{:.4f}_bleu_{:.4f}.pkl'.
format(epoch, reg_res["wer"], tran_res["bleu"]["bleu4"]))
trainer.save_checkpoint(save_ckpt, epoch, num_updates, loss)
# TODO? lr schedule.
model_manager.update(save_ckpt, reg_res["wer"], epoch)
# model_manager_bleu.update(save_ckpt, tran_res["bleu"]["bleu4"], epoch)
if opts.mode == "test":
# print("HERE!!!")
eval(opts, valid_datasets, trainer, epoch, gls_text_vocab)
eval(opts, test_datasets, trainer, epoch, gls_text_vocab)
exit()
def train(opts, train_datasets, trainer, epoch):
train_iter = trainer.get_batch_iterator(train_datasets, batch_size=opts.batch_size, shuffle=True,
num_workers=opts.num_workers)
epoch_loss, epoch_reg_loss, epoch_tran_loss = [], [], []
for samples in train_iter:
loss, reg_loss, tran_loss, num_updates = trainer.train_step(samples)
epoch_loss.append(loss.item())
epoch_reg_loss.append(reg_loss.item())
epoch_tran_loss.append(tran_loss.item())
lrs = trainer.get_lr()
if (num_updates % opts.print_step) == 0:
logging.info('Epoch: {:d}, num_updates: {:d}, loss: {:.3f}, reg_loss: {:.3f}, tran_loss: {:.3f}, '
'slr_lr: {:.6f}, slt_lr: {:.6f}'.
format(epoch, num_updates, loss, reg_loss, tran_loss, lrs["slr"], lrs["slt"]))
logging.info('Epoch: {:d}, loss: {:.3f}, reg_loss: {:.3f}, tran_loss: {:.3f}'.
format(epoch, np.mean(epoch_loss), np.mean(epoch_reg_loss), np.mean(epoch_tran_loss)))
return np.mean(epoch_loss)
def eval(opts, valid_datasets, trainer, epoch, gls_text_vocab):
eval_iter = trainer.get_batch_iterator(valid_datasets, batch_size=opts.batch_size, shuffle=False,
num_workers=opts.num_workers)
decoded_dict = {}
tran_outputs = []
val_err, val_correct, val_count = np.zeros([4]), 0, 0
for samples in tqdm(eval_iter):
err, correct, count, tran_outputs = trainer.valid_step(samples, decoded_dict, tran_outputs)
val_err += err
val_correct += correct
val_count += count
logging.info('-' * 50)
logging.info('Epoch: {:d}, DEV ACC: {:.5f}, {:d}/{:d}'.format(epoch, val_correct / val_count, val_correct, val_count))
logging.info('Epoch: {:d}, DEV WER: {:.5f}, SUB: {:.5f}, INS: {:.5f}, DEL: {:.5f}'.format(epoch,
val_err[0] / val_count, val_err[1] / val_count, val_err[2] / val_count, val_err[3] / val_count))
if opts.dataset_version == "phoenix_2014_trans":
gls_cln_fn = clean_phoenix_2014_trans
elif opts.dataset_version == "phoenix_2014":
gls_cln_fn = clean_phoenix_2014
else:
raise ValueError("Unknown Dataset Version: " + opts.dataset_version)
total_error = total_del = total_ins = total_sub = total_ref_len = 0
for vname, (ref, hyp) in decoded_dict.items():
ref_sent = gls_cln_fn(gls_text_vocab.gloss_lis_to_sentences(ref))
hyp_sent = gls_cln_fn(gls_text_vocab.gloss_lis_to_sentences(hyp))
res = wer_single(ref_sent, hyp_sent)
total_error += res["num_err"]
total_del += res["num_del"]
total_ins += res["num_ins"]
total_sub += res["num_sub"]
total_ref_len += res["num_ref"]
wer = (total_error / total_ref_len) * 100
del_rate = (total_del / total_ref_len) * 100
ins_rate = (total_ins / total_ref_len) * 100
sub_rate = (total_sub / total_ref_len) * 100
logging.info('Epoch: {:d}, DEV WER: {:.5f}, SUB: {:.5f}, INS: {:.5f}, DEL: {:.5f}'
.format(epoch, wer, sub_rate, ins_rate, del_rate))
# translation
txt_refs, txt_hyps = [], []
for i in range(len(tran_outputs)):
txt_hyp = gls_text_vocab.text_list_to_sentences(tran_outputs[i][0])
txt_ref = gls_text_vocab.text_list_to_sentences(tran_outputs[i][1])
# print("txt_hyp: ", txt_hyp)
# print("txt_ref: ", txt_ref)
txt_refs.append(txt_ref)
txt_hyps.append(txt_hyp)
txt_bleu = bleu(references=txt_refs, hypotheses=txt_hyps)
txt_chrf = chrf(references=txt_refs, hypotheses=txt_hyps)
txt_rouge = rouge(references=txt_refs, hypotheses=txt_hyps)
logging.info('Epoch: {:d}, BLEU-1: {:.5f}, BLEU-2: {:.5f}, BLEU-3: {:.4f}, BLEU-4: {:.4f}'.
format(epoch, txt_bleu["bleu1"], txt_bleu["bleu2"], txt_bleu["bleu3"], txt_bleu["bleu4"]))
logging.info('Epoch: {:d}, CHRF: {:.5f}, ROUGE: {:.5f}'
.format(epoch, txt_chrf, txt_rouge))
reg_res = {"wer": wer, "del_rate": del_rate, "ins_rate": ins_rate, "sub_rate": sub_rate,}
tran_res = {"bleu": txt_bleu, "chrf": txt_chrf, "rouge": txt_rouge}
return reg_res, tran_res
if __name__ == "__main__":
main()