-
Notifications
You must be signed in to change notification settings - Fork 20
/
train.py
275 lines (230 loc) · 8.99 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
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.utils.data
import models
import data.dataloader as dataloader
import data.utils as utils
import data.dict as dict
from optims import Optim
import lr_scheduler as L
import os
import argparse
import time
import json
import collections
import codecs
#config
parser = argparse.ArgumentParser(description='train.py')
parser.add_argument('-mask_train', default=False, type=bool,
help="mask_train")
parser.add_argument('-mask_sample', default=False, type=bool,
help="mask_sample")
parser.add_argument('-global_score', default=False, type=bool,
help="global_score")
parser.add_argument('-config', default='config.yaml', type=str,
help="config file")
parser.add_argument('-gpus', default=[1], type=int,
help="Use CUDA on the listed devices.")
parser.add_argument('-restore', default='./data/data/log/best_micro_f1_checkpoint.pt', type=str,
help="restore checkpoint")
parser.add_argument('-seed', type=int, default=1234,
help="Random seed")
parser.add_argument('-model', default='seq2seq', type=str,
help="Model selection")
parser.add_argument('-score', default='', type=str,
help="score_fn")
parser.add_argument('-pretrain', default=False, type=bool,
help="load pretrain embedding")
parser.add_argument('-notrain', default=False, type=bool,
help="train or not")
parser.add_argument('-limit', default=0, type=int,
help="data limit")
parser.add_argument('-log', default='', type=str,
help="log directory")
parser.add_argument('-unk', default=True, type=bool,
help="replace unk")
parser.add_argument('-memory', default=False, type=bool,
help="memory efficiency")
parser.add_argument('-label_dict_file', default='./data/data/topic_sorted.json', type=str,
help="label_dict")
opt = parser.parse_args()
config = utils.read_config(opt.config)
torch.manual_seed(opt.seed)
# checkpoint
if opt.restore:
print('loading checkpoint...\n')
checkpoints = torch.load(opt.restore)
# cuda
#use_cuda = torch.cuda.is_available() and len(opt.gpus) > 0
use_cuda = True
if use_cuda:
#torch.cuda.set_device(opt.gpus[0])
torch.cuda.manual_seed(opt.seed)
print(use_cuda)
# data
print('loading data...\n')
start_time = time.time()
datas = torch.load(config.data)
print('loading time cost: %.3f' % (time.time()-start_time))
trainset, validset = datas['train'], datas['valid']
src_vocab, tgt_vocab = datas['dicts']['src'], datas['dicts']['tgt']
config.src_vocab = src_vocab.size()
config.tgt_vocab = tgt_vocab.size()
trainloader = dataloader.get_loader(trainset, batch_size=config.batch_size, shuffle=True, num_workers=2)
validloader = dataloader.get_loader(validset, batch_size=config.batch_size, shuffle=False, num_workers=2)
if opt.pretrain:
pretrain_embed = torch.load(config.emb_file)
else:
pretrain_embed = None
# model
print('building model...\n')
model = getattr(models, opt.model)(config, src_vocab.size(), tgt_vocab.size(), use_cuda,
pretrain=pretrain_embed, score_fn=opt.score)
if opt.restore:
model.load_state_dict(checkpoints['model'])
if use_cuda:
model.cuda()
if len(opt.gpus) > 1:
model = nn.DataParallel(model, device_ids=opt.gpus, dim=1)
optim = Optim(config.optim, config.learning_rate, config.max_grad_norm,
lr_decay=config.learning_rate_decay, start_decay_at=config.start_decay_at)
optim.set_parameters(model.parameters())
if config.schedule:
scheduler = L.CosineAnnealingLR(optim.optimizer, T_max=config.epoch)
# total number of parameters
param_count = 0
for param in model.parameters():
param_count += param.view(-1).size()[0]
if not os.path.exists(config.log):
os.mkdir(config.log)
if opt.log == '':
log_path = config.log + utils.format_time(time.localtime()) + '/'
else:
log_path = config.log + opt.log + '/'
if not os.path.exists(log_path):
os.mkdir(log_path)
logging = utils.logging(log_path+'log.txt')
logging_csv = utils.logging_csv(log_path+'record.csv')
for k, v in config.items():
logging("%s:\t%s\n" % (str(k), str(v)))
logging("\n")
logging(repr(model)+"\n\n")
logging('total number of parameters: %d\n\n' % param_count)
logging('score function is %s\n\n' % opt.score)
if opt.restore:
updates = checkpoints['updates']
ori_updates = updates
else:
updates = 0
total_loss, start_time = 0, time.time()
report_total, report_correct = 0, 0
report_vocab, report_tot_vocab = 0, 0
scores = [[] for metric in config.metric]
scores = collections.OrderedDict(zip(config.metric, scores))
with open(opt.label_dict_file, 'r') as f:
label_dict = json.load(f)
# train
def train(epoch):
e = epoch
model.train()
if config.schedule:
scheduler.step()
print("Decaying learning rate to %g" % scheduler.get_lr()[0])
if opt.model == 'gated':
model.current_epoch = epoch
global e, updates, total_loss, start_time, report_total
score = eval(epoch)
model.train()
for raw_src, src, src_len, raw_tgt, tgt, tgt_len in trainloader:
src = Variable(src)
tgt = Variable(tgt)
src_len = Variable(src_len).unsqueeze(0)
tgt_len = Variable(tgt_len).unsqueeze(0)
if use_cuda:
src = src.cuda()
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
model.zero_grad()
loss = model.compute_reward(src, src_len, tgt, tgt_len)
loss.backward()
optim.step()
updates += 1
if updates % config.eval_interval == 0:
logging("time: %6.3f, epoch: %3d, updates: %8d, train loss: %6.3f\n"
% (time.time()-start_time, epoch, updates, loss))
print('evaluating after %d updates...\r' % updates)
score = eval(epoch)
for metric in config.metric:
scores[metric].append(score[metric])
if metric == 'macro_f1' and score[metric] >= max(scores[metric]):
save_model(log_path+'best_'+metric+'_checkpoint.pt')
if metric == 'micro_f1' and score[metric] >= max(scores[metric]):
save_model(log_path+'best_'+metric+'_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= max(scores[metric]):
save_model(log_path+'best_'+metric+'_checkpoint.pt')
model.train()
total_loss = 0
start_time = 0
report_total = 0
if updates % config.save_interval == 0:
save_model(log_path+'checkpoint.pt')
def eval(epoch):
model.eval()
reference, candidate, source, alignments = [], [], [], []
for raw_src, src, src_len, raw_tgt, tgt, tgt_len in validloader:
if len(opt.gpus) > 1:
samples, alignment = model.module.sample(src, src_len)
else:
samples, alignment = model.beam_sample(src, src_len, beam_size=config.beam_size)
candidate += [tgt_vocab.convertToLabels(s, dict.EOS) for s in samples]
source += raw_src
reference += raw_tgt
alignments += [align for align in alignment]
if opt.unk:
cands = []
for s, c, align in zip(source, candidate, alignments):
cand = []
for word, idx in zip(c, align):
if word == dict.UNK_WORD and idx < len(s):
try:
cand.append(s[idx])
except:
cand.append(word)
print("%d %d\n" % (len(s), idx))
else:
cand.append(word)
cands.append(cand)
candidate = cands
score = {}
result = utils.eval_metrics(reference, candidate, label_dict, log_path)
logging_csv([e, updates, result['hamming_loss'], \
result['macro_f1'], result['macro_precision'], result['macro_recall'],\
result['micro_f1'], result['micro_precision'], result['micro_recall']])
print('hamming_loss: %.8f | macro_f1: %.4f | micro_f1: %.4f'
% (result['hamming_loss'], result['macro_f1'], result['micro_f1']))
#score['one_error'] = result['one_error']
score['hamming_loss'] = result['hamming_loss']
score['macro_f1'] = result['macro_f1']
score['micro_f1'] = result['micro_f1']
return score
def save_model(path):
global updates
model_state_dict = model.module.state_dict() if len(opt.gpus) > 1 else model.state_dict()
checkpoints = {
'model': model_state_dict,
'config': config,
'optim': optim,
'updates': updates}
torch.save(checkpoints, path)
def main():
for i in range(1, config.epoch+1):
if not opt.notrain:
train(i)
else:
eval(i)
for metric in config.metric:
logging("Best %s score: %.2f\n" % (metric, max(scores[metric])))
if __name__ == '__main__':
main()