-
Notifications
You must be signed in to change notification settings - Fork 0
/
run.py
340 lines (299 loc) · 12 KB
/
run.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
import os
import torch
import random
# from apex import amp
random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
import torch.nn as nn
import numpy as np
np.random.seed(0)
torch.backends.cudnn.deterministic = True
import tqdm
import argparse
import config_utils
from dataset_utils import (
load_vocab_new,
tokenize,
bpe_tokenize,
)
# from model import DualTransformer
# from bert_model import DualTransformer
from model_adapter import DualTransformer
from torch.utils import data
from torch import optim
from bpemb import BPEmb
import math
import time
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
def len_to_mask(len_seq, max_len=None):
"""len to mask"""
len_seq = len_seq.int()
if max_len is None:
max_len = torch.max(len_seq).item()
mask = torch.zeros((len_seq.size(0), max_len))
for i, l in enumerate(len_seq):
mask[i, :l] = 1
return mask
def train(model, data_loader, optimizer, n_gpu, FLAGS):
model.train()
train_loss = 0.0
train_step = 0.0
train_right, train_total = 0.0, 0.0
# for i, _data in enumerate(data_loader):
for i, _data in tqdm.tqdm(enumerate(data_loader)):
sv, sv_len, cv, cv_len, rv, tv, tv_len, tstr = _data
batch = {}
batch["tgt_str"] = tstr
batch["src"] = sv
batch["src_len"] = sv_len
batch["src_mask"] = len_to_mask(sv_len)
batch["con"] = cv
batch["con_len"] = cv_len
batch["con_mask"] = len_to_mask(cv_len)
batch["rel"] = rv
# batch["word_rel_mask"] = word_rel_mask
# batch["wr"] = wr
batch["tgt_input"] = tv[:, :-1] # remove eos
batch["tgt_ref"] = tv[:, 1:] # remove bos
batch["tgt_len"] = tv_len - 1
batch["tgt_mask"] = len_to_mask(tv_len - 1)
outputs = model(batch)
loss = outputs["loss"]
if n_gpu > 1:
loss = loss.mean()
if FLAGS.grad_accum_steps > 1:
loss = loss / FLAGS.grad_accum_steps
if FLAGS.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward() # just calculate gradient
if i % FLAGS.grad_accum_steps == 0: # optimizer step
optimizer.step()
optimizer.zero_grad()
train_loss += float(loss.cpu().item())
train_step += 1.0
train_right += outputs["counts"][0].mean().cpu().item()
train_total += outputs["counts"][1].mean().cpu().item()
torch.cuda.empty_cache()
return train_loss / train_step, train_right / train_total
def validate(model, data_loader, n_gpu):
model.eval()
with torch.no_grad():
dev_loss = 0.0
dev_step = 0.0
dev_right, dev_total = 0.0, 0.0
for i, _data in tqdm.tqdm(enumerate(data_loader)):
# sv, sv_len, cv, cv_len, rv, tv, tv_len, tstr = _data
sv, sv_len, cv, cv_len, rv, tv, tv_len, tstr = _data
batch = {}
batch["tgt_str"] = tstr
batch["src"] = sv
batch["src_len"] = sv_len
batch["src_mask"] = len_to_mask(sv_len)
batch["con"] = cv
batch["con_len"] = cv_len
batch["con_mask"] = len_to_mask(cv_len)
batch["rel"] = rv
# batch["word_rel_mask"] = word_rel_mask
# batch["wr"] = wr
batch["tgt_input"] = tv[:, :-1] # remove eos
batch["tgt_ref"] = tv[:, 1:] # remove bos
batch["tgt_len"] = (tv_len - 1)
batch["tgt_mask"] = len_to_mask(tv_len - 1)
outputs = model(batch)
loss = outputs["loss"]
if n_gpu > 1:
loss = loss.mean()
dev_loss += float(loss.cpu().item())
dev_step += 1.0
dev_right += outputs["counts"][0].mean().cpu().item()
dev_total += outputs["counts"][1].mean().cpu().item()
torch.cuda.empty_cache()
return dev_loss / dev_step, dev_right / dev_total
if __name__ == "__main__":
argparser = argparse.ArgumentParser()
argparser.add_argument("--config_path", type=str, help="Configuration file.")
FLAGS, unparsed = argparser.parse_known_args()
if FLAGS.config_path is not None:
print("Loading hyperparameters from " + FLAGS.config_path)
FLAGS = config_utils.load_config(FLAGS.config_path)
log_dir = FLAGS.log_dir
continue_train = False
if not os.path.exists(log_dir):
os.makedirs(log_dir)
else:
continue_train = True
path_prefix = log_dir + "/wiki.{}".format(FLAGS.suffix)
log_file = open(path_prefix + ".log", "w+")
if not continue_train:
log_file.write("{}\n".format(str(FLAGS)))
log_file.flush()
print("Log file path: {}".format(path_prefix + ".log"))
config_utils.save_config(FLAGS, path_prefix + ".config.json")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
# n_gpu = 0
print(
"device: {}, n_gpu: {}, grad_accum_steps: {}".format(device, n_gpu, FLAGS.grad_accum_steps)
)
log_file.write(
"device: {}, n_gpu: {}, grad_accum_steps: {}\n".format(
device, n_gpu, FLAGS.grad_accum_steps
)
if not continue_train
else ""
)
# exit()
s_time = time.time()
words, word2id, concepts, concept2id, relations, relation2id= load_vocab_new(FLAGS.save_data)
print("Loading vocab takes {:.3f}s".format(time.time() - s_time))
print("Vocabulary size: {}".format(len(words)))
if FLAGS.save_data != "":
tokenize_fn = tokenize
word_vocab, concept_vocab, relation_vocab = word2id, concept2id, relation2id,
else:
print("Not support other vocabulary for now!!")
exit()
checkpoint = None
best_checkpoint_path = path_prefix + "_best.checkpoint.bin"
last_checkpoint_path = path_prefix + "_last.checkpoint.bin"
if os.path.exists(last_checkpoint_path):
print("!!Existing checkpoint. Loading...")
log_file.write("!!Existing checkpoint. Loading...\n")
checkpoint = torch.load(last_checkpoint_path)
if FLAGS.use_bpe_pretrain:
# word_emb = torch.load(FLAGS.save_data + "/word_emb.pt")
word_emb = None
con_emb = torch.load(FLAGS.save_data + "/concept_emb.pt")
else:
word_emb = None
con_emb = None
model = DualTransformer(FLAGS, word_emb, con_emb, word2id, concept2id, relation2id)
model.to(device)
# model = DualTransformer(FLAGS, None, None, word2id, None, None)
optimizer = optim.Adam(model.parameters(), lr=FLAGS.learning_rate)
if FLAGS.fp16:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1") # 这里是“欧一”,不是“零一”
print(model)
print(
"num. model params: {} (num. trained: {})".format(
sum(p.numel() for p in model.parameters()),
sum(p.numel() for p in model.parameters() if p.requires_grad),
)
)
if not continue_train:
log_file.write(str(model))
log_file.write(
"num. model params: {} (num. trained: {})".format(
sum(p.numel() for p in model.parameters()),
sum(p.numel() for p in model.parameters() if p.requires_grad),
)
)
if n_gpu > 1:
model = nn.DataParallel(model)
if checkpoint:
if n_gpu <= 1:
new_pre = {}
for k, v in checkpoint["model_state_dict"].items():
name = k[7:] if k.startswith("module") else k
new_pre[name] = v
model.load_state_dict(new_pre)
else:
model.load_state_dict(checkpoint["model_state_dict"])
best_accu = 0.0
if checkpoint:
assert "best_accu" in checkpoint
best_accu = checkpoint["best_accu"]
print("Initial accuracy {:.4f}".format(best_accu))
log_file.write("Initial accuracy {:.4f}\n".format(best_accu))
if checkpoint:
optimizer.load_state_dict(checkpoint["optimizer"])
for state in optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.to(device)
start_epoch = checkpoint["epoch"]
else:
start_epoch = 0
# for the usage of BertAdam
# named_params = list(model.named_parameters())
# no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
# grouped_params = [
# {'params': [p for n, p in named_params if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
# {'params': [p for n, p in named_params if any(nd in n for nd in no_decay)], 'weight_decay': 0.01}
# ]
worker_init_fn = lambda worker_id: np.random.seed(np.random.get_state()[1][0] + worker_id)
print("Loading train data and making batches")
log_file.write("Loading data and making batches\n")
s_time = time.time()
train_set = torch.load(FLAGS.save_data + "/train_data.pt")
train_set.max_tok_len = 512
print("Loading data takes {:.3f}s".format(time.time() - s_time))
s_time = time.time()
train_loader = train_set.GetDataloader(
batch_size=FLAGS.batch_size, shuffle=FLAGS.is_shuffle, num_workers=4
)
print("Loading dev data and making batches")
dev_set = torch.load(FLAGS.save_data + "/test_data.pt")
dev_loader = dev_set.GetDataloader(batch_size=FLAGS.batch_size, shuffle=False, num_workers=1)
print("Loading dev data takes {:.3f}s".format(time.time() - s_time))
print("Num training examples = {}".format(len(train_set.instance)))
log_file.write("Num training examples = {}\n".format(len(train_set.instance)))
max_patience = FLAGS.patience
patience = 0
optimizer.zero_grad()
for iter in range(start_epoch, FLAGS.num_epochs):
train_loss, train_accu = train(model, train_loader, optimizer, n_gpu, FLAGS)
val_loss, val_accu = validate(model, dev_loader, n_gpu)
print(
"iter: {}, lr:{:.5f} TRAIN loss: {:.4f} accu: {:.4f}; VAL loss: {:.4f} accu: {:.4f}".format(
iter, optimizer.param_groups[0]["lr"], train_loss, train_accu, val_loss, val_accu
)
)
log_file.write(
"iter: {}, lr:{:.5f} TRAIN loss: {:.4f} accu: {:.4f}; VAL loss: {:.4f} accu: {:.4f}\n".format(
iter, optimizer.param_groups[0]["lr"], train_loss, train_accu, val_loss, val_accu
)
)
state = {
"epoch": iter,
"model_state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
"best_accu": best_accu,
}
# print("saving last model ...")
torch.save(state, last_checkpoint_path)
if best_accu < val_accu:
best_accu = val_accu
patience = 0
# save model
print("saving best model ...")
log_file.write("saving best model ...\n")
config_utils.save_config(FLAGS, path_prefix + ".config.json")
state = {
"epoch": iter,
"model_state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
"best_accu": best_accu,
}
torch.save(state, best_checkpoint_path)
os.system("cp {} {}".format(best_checkpoint_path, last_checkpoint_path))
else:
patience += 1
if patience >= max_patience:
print("Reaching max patience! exit...")
exit()
if iter % 10 == 0 and iter > 0:
# state = {
# "epoch": iter,
# "model_state_dict": model.state_dict(),
# "optimizer": optimizer.state_dict(),
# "best_accu": best_accu,
# }
# print("saving last model ...")
tmp_checkpoint_path = path_prefix + "_epoch_{}.checkpoint.bin".format(iter)
# os.system("cp {} {}".format(last_checkpoint_path, tmp_checkpoint_path))
os.system("cp {} {}.{}".format(best_checkpoint_path, best_checkpoint_path, iter))
# torch.cuda.empty_cache()