-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
201 lines (167 loc) · 7.65 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
import argparse
import datetime
import time
import shutil
from tqdm import tqdm
from torch import optim
import torch.nn as nn
from utils import *
from data_processor import DataProcessor
from evals import BleuScore, RougeScore
from early_stopping import EarlyStopping
from model import EncoderAttn, DecoderAttn, Table2Text
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', default=32,
type=int, help='Mini batch size')
parser.add_argument('--max_epoch', default=40, type=int,
help='Max epoch for training.')
parser.add_argument('--lr', default=3e-4, type=float,
help='Initial learning rate.')
parser.add_argument('--field_emb_dim', default=50, type=int,
help='Dimension of field embedding.')
parser.add_argument('--pos_emb_dim', default=5, type=int,
help='Dimension of position embedding.')
parser.add_argument('--word_emb_dim', default=400, type=int,
help='Dimension of word embedding.')
parser.add_argument('--hidden_dim', default=500, type=int,
help='Dimension of hidden layer.')
parser.add_argument('--dropout', default=0.3, type=float, help='Dropout rate.')
parser.add_argument('--weight_decay', default=0.0,
type=float, help='Weight decay (L2 penalty).')
parser.add_argument('--grad_clip', default=5.0, type=float,
help='Max norm of the gradients clipping.')
parser.add_argument('--random_seed', default=1, type=int,
help='Seed for generating random numbers.')
parser.add_argument('--beam_width', default=1, type=int,
help='Size of beam search width.')
parser.add_argument('--max_len', default=60, type=int,
help='Max length of the texts.')
parser.add_argument('--max_field', default=100, type=int,
help='Max length of the fields.')
parser.add_argument('--pos_size', default=31, type=int,
help='Max number of position.')
parser.add_argument('--train', default=False, type=bool,
help='If False, then doing inference.')
parser.add_argument('--resume', default=True, type=bool,
help='Whether to load checkpoints to resume training.')
parser.add_argument('--copy', default=True, type=bool,
help='Whether to use copy mechanism.')
args = parser.parse_args()
fix_seed(args.random_seed)
cur_time = datetime.datetime.now().strftime('%Y_%m_%d_%H_%M')
logger = get_logger('./results/logs/' + cur_time + '.log')
for arg in vars(args): # print arguments
logger.info("{} : {}".format(arg, getattr(args, arg)))
start_time = time.time()
data_processor = DataProcessor(args)
train_data_loader = data_processor.get_data_loader(
mode='train', batch_size=args.batch_size, shuffle=True)
dev_data_loader = data_processor.get_data_loader(
mode='dev', batch_size=args.batch_size, shuffle=False)
field_vocab_size = data_processor.field_vocab_size
pos_size = data_processor.pos_size
word_vocab_size = data_processor.word_vocab_size
logger.info(f'data processing consumes: {(time.time() - start_time):.2f}s')
def weights_init(m: nn.Module):
for name, param in m.named_parameters():
if 'weight' in name:
nn.init.normal_(param.data, mean=0.0, std=0.01)
else:
nn.init.constant_(param.data, 0)
encoder = EncoderAttn(field_vocab_size, pos_size, word_vocab_size,
args.field_emb_dim, args.pos_emb_dim, args.word_emb_dim,
args.hidden_dim, args.dropout)
decoder = DecoderAttn(word_vocab_size,
args.word_emb_dim, args.hidden_dim, args.dropout)
model = Table2Text(encoder, decoder, args.beam_width,
args.max_len, args.max_field).to(device)
optimizer = optim.Adam(
model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
criterion = nn.NLLLoss()
early_stop = EarlyStopping(mode='min', min_delta=0.001, patience=5)
if args.resume:
checkpoint, cp_name = load_checkpoint(latest=True)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
logger.info(f'load checkpoint: [{cp_name}]')
def train(data_loader):
model.train()
epoch_loss = 0
for train_input, train_target in tqdm(data_loader):
optimizer.zero_grad()
train_output = model(train_input, train_target, train_mode=True)
train_output = train_output[:, :-1].reshape(-1, train_output.size(-1))
loss = criterion(train_output, train_target[:, 1:].reshape(-1))
loss.backward()
epoch_loss += loss.item()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
return epoch_loss / len(data_loader)
def validate(data_loader):
model.eval()
epoch_loss = 0
with torch.no_grad():
for validate_input, validate_target in data_loader:
validate_output = model(
validate_input, validate_target, train_mode=True)
validate_output = validate_output[:, :-
1].reshape(-1, validate_output.size(-1))
loss = criterion(
validate_output, validate_target[:, 1:].reshape(-1))
epoch_loss += loss.item()
return epoch_loss / len(data_loader)
def evaluate(infer_input):
model.eval()
with torch.no_grad():
infer_target = torch.tensor(
[2], dtype=torch.long, device=device) # SOS_TOKEN
eval_output, attn = model(infer_input, infer_target, train_mode=False)
return eval_output, attn
loss_dict_train, loss_dict_dev = [], []
for epoch in range(1, int(args.max_epoch + 1)):
start_time = time.time()
train_loss = train(train_data_loader)
dev_loss = validate(dev_data_loader)
loss_dict_train.append(train_loss)
loss_dict_dev.append(dev_loss)
epoch_min, epoch_sec = record_time(start_time, time.time())
logger.info(
f'epoch: [{epoch:02}/{args.max_epoch}] train_loss={train_loss:.3f} valid_loss={dev_loss:.3f} '
f'duration: {epoch_min}m {epoch_sec}s')
if early_stop.step(dev_loss):
logger.info(f'early stop at [{epoch:02}/{args.max_epoch}]')
break
if args.max_epoch > 0:
save_checkpoint(experiment_time=cur_time, model=model, optimizer=optimizer)
check_file_exist('./results/rouge/system')
check_file_exist('./results/rouge/gold')
bleu_scorer = BleuScore()
rouge_scorer = RougeScore(system_dir='./results/rouge/system',
model_dir='./results/rouge/gold',
n_gram=4)
ref_summaries = data_processor.get_refs(tag='test')
bleu_scorer.set_refs(ref_summaries)
rouge_scorer.set_refs(ref_summaries)
for idx_data in range(len(data_processor.test_data)):
seq_input = torch.tensor(data_processor.process_one_data(
idx_data=idx_data), dtype=torch.long, device=device).unsqueeze(0)
seq_output, attn_map = evaluate(seq_input)
list_seq = seq_output.squeeze().tolist()
if not args.copy:
text_gen = data_processor.translate(list_seq)
else:
text_gen = data_processor.translate_w_copy(
list_seq=list_seq, attn_score=attn_map, data_idx=idx_data)
bleu_scorer.add_gen(text_gen)
rouge_scorer.add_gen(text_gen)
bleu_score = bleu_scorer.calculate()
logger.info(f'bleu score: {bleu_score:.2f}')
rouge_scorer.file_writer()
rouge_score_dict = rouge_scorer.calculate()
for n_gram in rouge_score_dict:
logger.info(f'{n_gram}: {rouge_score_dict[n_gram]:.2f}')
clean_logs(latest=True) # clean chaotic logs caused by pyrouge
# remove redundant folders
shutil.rmtree('./results/rouge/system')
shutil.rmtree('./results/rouge/gold')