-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
297 lines (286 loc) · 16.8 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
import argparse
import pickle
import os
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from src.basic.add_args import add_basic_arguments, add_boundary_arguments, add_optimizer_arguments
from src.basic.constant import POS_TO_IDX, NER_TO_IDX
from src.basic.utils import make_cache_files, config_logs
from src.models.boundary import BoundaryEncoderDecoder, HybridBoundaryEncoderDecoder
from src.models.encdec import EncoderDecoder, AttentionEncoderDecoder, BahdanauAttnEncoderDecoder, \
FeatureRicherEncoderDecoder
from src.models.hierarchical import BasicHierarchicalEncoderDecoder, AttentionHierarchicalEncoderDecoder
from src.models.intra_attn_seq2seq import IntraEncoderDecoder
from src.models.evaluator import Evaluator
from src.models.trainer import Trainer
from src.preprocess.vocab import Vocabulary
from src.preprocess.preprocess import AMIDataset, CNNDataset
from src.models.optimizer import Optim
def train(dataset, model, vocab, args, logger, pre_model=None):
logger.info(model)
input_vocab = vocab["input_vocab"]
output_vocab = vocab["input_vocab"]
if args.model in ["hierarchical-encdec", "attn-hierarchical-encdec"]:
train_iter = dataset.generate(batch_size=args.batch_size, input_vocab=input_vocab, args=args,
output_vocab=output_vocab, batch_type="sentence", _type="train", early_stop=None)
valid_iter = dataset.generate(batch_size=args.batch_size, input_vocab=input_vocab, args=args,
output_vocab=output_vocab, batch_type="sentence", _type="valid", early_stop=None)
test_iter = dataset.generate(batch_size=args.batch_size, input_vocab=input_vocab, args=args,
output_vocab=output_vocab, batch_type="sentence", _type="valid", early_stop=None)
else:
train_iter = dataset.generate(batch_size=args.batch_size, input_vocab=input_vocab, args=args,
output_vocab=output_vocab, _type="train", early_stop=None)
valid_iter = dataset.generate(batch_size=args.batch_size, input_vocab=input_vocab, args=args,
output_vocab=output_vocab, _type="test", early_stop=None)
test_iter = dataset.generate(batch_size=1, input_vocab=input_vocab, args=args,
output_vocab=output_vocab, _type="test", early_stop=None)
num_in_total_test = next(test_iter)
evaluator = Evaluator(vocab, model, logger, args)
optimizer = Optim(
args.optim, args.lr, args.max_grad_norm,
lr_decay=args.learning_rate_decay,
start_decay_at=args.start_decay_at,
beta1=args.adam_beta1,
beta2=args.adam_beta2,
adagrad_accum=args.adagrad_accumulator_init,
opt=args
)
optimizer.set_parameters(model.parameters())
trainer = Trainer(model=model, vocab=vocab, train_iter=train_iter, valid_iter=valid_iter,
train_loss=nn.NLLLoss(size_average=False), valid_loss=nn.NLLLoss(size_average=False),
optim=optimizer, logger=logger, args=args, pre_model=pre_model)
best_ppl = np.inf
best_bleu_output = ""
best_rouge_output = ""
best_rouge = 0
for epoch in range(args.epoch_num):
logger.info('\n')
# 1. Train for one epoch on the training set.
train_stats = trainer.train(epoch, print_every=args.print_every)
logger.info('Train loss: %g' % train_stats.get_loss())
logger.info('Train perplexity: %g' % train_stats.ppl())
logger.info('Train accuracy: %g' % train_stats.accuracy())
# 2. Validate on the validation set.
# valid_stats = trainer.validate()
# logger.info('Validation loss: %g' % valid_stats.get_loss())
# logger.info('Validation perplexity: %g' % valid_stats.ppl())
# logger.info('Validation accuracy: %g' % valid_stats.accuracy())
evaluator.model = trainer.model
res = evaluator.test(num_in_total=num_in_total_test, test_iter=test_iter)
# 3. dump the checkpoints
ckp = trainer.save_checkpoint(args, model, epoch, train_stats)
# if args.reinforce:
r = res["rouge_su4"]
if r > best_rouge:
logger.info("It's best model")
best_bleu_output = res["bleu_output"]
best_rouge_output = res["rouge_output"]
best_ckp = ckp
best_rouge = r
torch.save(best_ckp, "models/{}/best-model.pt".format(
args.model + "-" + args.dataset + "-" + args.path_prefix))
# else:
# if valid_stats.ppl() < best_ppl:
# best_ckp = ckp
# best_ppl = valid_stats.ppl()
# torch.save(best_ckp, "models/{}/best-model.pt".format(
# args.model + "-" + args.dataset + "-" + args.path_prefix))
# 4. Update the learning rate
trainer.epoch_step(r, epoch)
print("\nBest Bleu: ", best_bleu_output)
print("Best Rouge: ", best_rouge_output)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
add_basic_arguments(parser)
add_boundary_arguments(parser)
add_optimizer_arguments(parser)
# args = parser.parse_args(["--model", "attn-encdec", # "b-attn-encdec" "attn-encdec" "encdec"
# "--dataset", "ami",
# "--load_model", "models/attn-encdec-ami/best-model.pt",
# "--epoch-num", "10",
# "--batch-size", "5",
# "--hidden-size", "512",
# "--word-embed-size", "256",
# "--dropout", "0.3",
# "--beam-size", "1",
# "--print-every", "5",
# "--max-sent-num", "8",
# "--max-sent-len", "20",
# "--use-cuda",
# # "--test"
# ])
# Flatten Model Args
# args = parser.parse_args(["--data-path", "/home/panhaojie/AbsDialogueSum/checkpoints/cpt-1/data",
# "--model", "hybrid-boundary-encdec",
# "--dataset", "ami",
# # "--load_model", "models/hybrid-boundary-encdec-ami-h-512-e-196/model_acc_32.40_ppl_86.28_e11.pt",
# # "--load_model", "checkpoints/cpt-1/models/hybrid-fr.pt",
# # "--load_model", "models/hybrid-boundary-encdec-ami-reinforce/best-model.pt",
# # "--load_model", "checkpoints/alpha/ml-rl-9988.pt",
# "--load_model", "models/hybrid-boundary-encdec-ami-FIX-50/model_acc_47.65_ppl_12.11_e6.pt",
# # "--load_w2v", "glove.6B.100d.pt",
# "--epoch-num", "15",
# "--batch-size", "1",
# "--hidden-size", "512",
# "--word-embed-size", "256",
# "--dropout", "0.5",
# "--beam-size", "1",
# "--print-every", "20",
# "--encoder-max-len", "6000",
# "--decoder-max-len", "160",
# "--use-cuda",
# "--path-prefix", "FIX-50",
# "--feature_rich",
# "--n_layers", "1",
# # "--reinforce",
# "-max_grad_norm", "5",
# "-lr", "1",
# '-learning_rate_decay', "0.5",
# "--alpha", "0.9985",
# # "--test"
# ])
args = parser.parse_args()
data_root = args.data_path
if args.model in ["hierarchical-encdec", "attn-hierarchical-encdec"]:
with open(os.path.join(data_root, "%s-vocab-hierachy.pkl" % args.dataset), "rb") as f:
vocab = pickle.load(f)
else:
with open(os.path.join(data_root, "%s-vocab.pkl" % args.dataset), "rb") as f:
vocab = pickle.load(f)
if not os.path.exists(os.path.join(data_root, "%s-dataset.pkl" % args.dataset)):
print("Create Dataset %s" % args.dataset)
if args.dataset == "ami":
dataset = AMIDataset(data_root)
elif args.dataset == "cnn":
dataset = CNNDataset(os.path.join(data_root, "cnn-data"))
with open(os.path.join(data_root, "%s-dataset.pkl" % args.dataset), "wb") as f:
pickle.dump(dataset, f)
else:
print("Load Dataset %s" % args.dataset)
with open(os.path.join(data_root, "%s-dataset.pkl" % args.dataset), "rb") as f:
dataset = pickle.load(f)
make_cache_files(dataset, args)
logger = config_logs(args)
logger.info(args)
if args.test:
logger.info(args.load_model)
logger.info("=" * 10 + " Test " + "=" * 10)
if args.model not in ["textRank"]:
ckp = torch.load(args.load_model)
model = ckp["model"]
else:
model = None
logger.info(model)
logger.info("BEAM SIZE: {}".format(args.beam_size))
evaluator = Evaluator(vocab, model, logger, args)
if args.model == "hierarchical-encdec":
test_iter = dataset.generate(batch_size=1, input_vocab=vocab["input_vocab"], args=args,
output_vocab=vocab["output_vocab"], batch_type="sentence", _type="test",
early_stop=None)
else:
test_iter = dataset.generate(batch_size=1, input_vocab=vocab["input_vocab"],
output_vocab=vocab["output_vocab"], args=args, _type="test")
num_in_total = next(test_iter)
evaluator.test(num_in_total=num_in_total, test_iter=test_iter)
else:
pretrained_model = None
tmp_vocab = vocab["output_vocab"] if args.model == "intra-attn" else vocab["input_vocab"]
if args.model == "encdec":
logger.info("Basic Encoder Decoder")
model = EncoderDecoder(encoder_vocab_size=vocab["input_vocab"].size,
decoder_vocab_size=tmp_vocab.size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
rnn_type=args.rnn_type, dropout=args.dropout, use_cuda=args.use_cuda)
model.init_params()
if args.load_w2v is not None:
pre_embedding = torch.load(os.path.join(data_root, args.load_w2v))["embed"]
model.load_word_vectors(pre_embedding)
elif args.model == "attn-encdec":
logger.info("Luong Attention Based Encoder Decoder")
model = AttentionEncoderDecoder(encoder_vocab_size=vocab["input_vocab"].size,
decoder_vocab_size=tmp_vocab.size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
n_layers=args.n_layers,
rnn_type=args.rnn_type, dropout=args.dropout, use_cuda=args.use_cuda)
if args.reinforce:
ckp = torch.load(args.load_model)
pretrained_model = ckp["model"]
model.init_params(pretrained_model)
elif args.model == "b-attn-encdec":
logger.info("Bahdanau Attention Based Encoder Decoder")
model = BahdanauAttnEncoderDecoder(encoder_vocab_size=vocab["input_vocab"].size,
decoder_vocab_size=tmp_vocab.size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
rnn_type=args.rnn_type, dropout=args.dropout, use_cuda=args.use_cuda)
elif args.model == "intra-attn":
logger.info("Intra Attention Based Encoder Decoder")
model = IntraEncoderDecoder(encoder_vocab_size=vocab["input_vocab"].size,
decoder_vocab_size=vocab["input_vocab"].size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
rnn_type=args.rnn_type, dropout=args.dropout, use_cuda=args.use_cuda)
elif args.model == "hierarchical-encdec":
logger.info("Basic Hierarchical Encoder Decoder")
model = BasicHierarchicalEncoderDecoder(vocab_size=vocab["input_vocab"].size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
rnn_type=args.rnn_type, dropout=args.dropout, use_cuda=args.use_cuda)
elif args.model == "attn-hierarchical-encdec":
logger.info("Attention Hierarchical Encoder Decoder")
model = AttentionHierarchicalEncoderDecoder(vocab_size=vocab["input_vocab"].size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
rnn_type=args.rnn_type, dropout=args.dropout,
use_cuda=args.use_cuda)
elif args.model == "boundary-encdec":
logger.info("Boundary Based Encoder Decoder")
if args.feature_rich:
extra_categorical_nums = [len(POS_TO_IDX), len(NER_TO_IDX), 11, 10]
extra_embed_sizes = [25, 5, 10, 10]
else:
extra_categorical_nums = None
extra_embed_sizes = None
model = BoundaryEncoderDecoder(encoder_vocab_size=vocab["input_vocab"].size,
decoder_vocab_size=vocab["input_vocab"].size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
extra_categorical_nums=extra_categorical_nums,
extra_embed_sizes=extra_embed_sizes,
bd_mid_size=args.bd_mid_size, n_layers=1, dropout=0.5,
use_cuda=args.use_cuda)
model.reset_parameters()
elif args.model == "hybrid-boundary-encdec":
logger.info("Hybrid Boundary Based Encoder Decoder")
if args.feature_rich:
extra_categorical_nums = [len(POS_TO_IDX), len(NER_TO_IDX), 11, 10]
extra_embed_sizes = [25, 5, 10, 10]
else:
extra_categorical_nums = None
extra_embed_sizes = None
model = HybridBoundaryEncoderDecoder(encoder_vocab_size=vocab["input_vocab"].size,
decoder_vocab_size=vocab["input_vocab"].size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
extra_categorical_nums=extra_categorical_nums,
extra_embed_sizes=extra_embed_sizes,
bd_mid_size=args.bd_mid_size,
use_cuda=args.use_cuda)
if args.reinforce:
ckp = torch.load(args.load_model)
pretrained_model = ckp["model"]
model.init_params(pretrained_model)
if args.load_model:
ckp = torch.load(args.load_model)
pretrained_model = ckp["model"]
model.init_params(pretrained_model)
elif args.model == "fr-encdec":
logger.info("Feature Rich Encoder Decoder")
model = FeatureRicherEncoderDecoder(encoder_vocab_size=vocab["input_vocab"].size,
decoder_vocab_size=tmp_vocab.size,
embed_size=args.word_embed_size, hidden_size=args.hidden_size,
extra_categorical_nums=[len(POS_TO_IDX), len(NER_TO_IDX), 11, 10],
extra_embed_sizes=[25, 5, 10, 10], n_layers=args.n_layers,
rnn_type=args.rnn_type, dropout=args.dropout, use_cuda=args.use_cuda)
if args.use_cuda:
model = model.cuda()
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
print("Parm Counter:", count_parameters(model))
train(dataset, model, vocab, args, logger, pretrained_model)