-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrain.py
155 lines (143 loc) · 6.96 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
#-*-coding:utf-8-*-
import torch
import numpy as np
import random
import time
import os
from util.processdataall import dataPreprocess, vocab, globalVocab
from util.makedata import build_pretrain_embedding, buildDataperBatch
from util.evaluate import evaluate
from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter
import gc
from model.sama import sama
from tqdm import tqdm
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
# remove GPU cache
torch.backends.cudnn.deterministic = True
setup_seed(421)
time_above = 0
def parse_args():
parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter, conflict_handler="resolve")
parser.add_argument("--file_dir", type=str, default="/home/rleating/datasets/SAMA")
parser.add_argument("--train", type=str, default="train.pt")
parser.add_argument("--test", type=str, default="test.pt")
parser.add_argument("--vocab", type=str, default="vocab.pt")
parser.add_argument("--model_dir", help="the model dir", type=str, default="./trained_model")
parser.add_argument("--result_dir", help="the result dir", type=str, default="./results/training")
parser.add_argument("--pretrain_tgt_embedding", type=str, default="pretrained_w2v")
parser.add_argument("--pretrain_src_embedding", type=str, default="pretrained_w2v")
parser.add_argument("--ifGPU", help="whether use gpu", type=bool, default=True)
parser.add_argument("--encoder_bidirectional", type=bool, default=True)
parser.add_argument("--encoder_layer", type=int, default=1)
parser.add_argument("--max_decoder_len", type=int, default=150)
parser.add_argument("--dropout", help="dropout rate", type=float, default=0.3)
parser.add_argument("--learning_rate", type=float, default=0.001)
parser.add_argument("--epoch", type=int, default=30)
parser.add_argument("--batch", type=int, default=5)
parser.add_argument("--device", type=int, default=0)
parser.add_argument("--src_embedding_dim", type=int, default=100)
parser.add_argument("--tgt_embedding_dim", type=int, default=100)
parser.add_argument("--src_encoder", type=str, default="LSTM")
parser.add_argument("--src_decoder", type=str, default="LSTM")
parser.add_argument("--tgt_encoder", type=str, default="LSTM")
parser.add_argument("--tgt_decoder", type=str, default="LSTM")
parser.add_argument("--src_encoder_hidden_dim", type=int, default=400)
parser.add_argument("--generatealpha", type=float, default=1.4)
parser.add_argument("--topicGenLamda", type=float, default=0.05)
parser.add_argument("--max_skill_len", type=int, default=30)
parser.add_argument("--skill_len", type=int, default=500)
parser.add_argument("--device", type=int, default=0)
args = parser.parse_args()
return args
if __name__ == "__main__":
# read config
data = parse_args()
if not os.path.exists(data.model_dir):
os.makedirs(data.model_dir)
if not os.path.exists(data.result_dir):
os.makedirs(data.result_dir)
# load vocb
vocab_dir = os.path.join(data.file_dir, data.vocab)
data.gVocab = torch.load(vocab_dir)
# load data
train, test = os.path.join(data.file_dir, data.train), os.path.join(data.file_dir, data.test)
data.train_dataset = torch.load(train)
data.test_dataset = torch.load(test)
# load pretrain embedding
pretrain_src_path = os.path.join(data.file_dir, data.pretrain_src_embedding)
data.pretrain_src_embedding, data.src_embedding_dim, ukn_src_count = build_pretrain_embedding(
pretrain_src_path, data.gVocab, "src")
src_embedding = torch.tensor(data.pretrain_src_embedding)
print("src unknown words: " + str(ukn_src_count))
pretrain_tgt_path = os.path.join(data.file_dir, data.pretrain_tgt_embedding)
data.pretrain_tgt_embedding, data.tgt_embedding_dim, ukn_tgt_count = build_pretrain_embedding(
pretrain_tgt_path, data.gVocab, "tgt")
tgt_embedding = torch.tensor(data.pretrain_tgt_embedding)
print("tgt unknown words: " + str(ukn_tgt_count))
# build model
model = sama(data)
param_count = 0 # counting the parameters
for param in model.parameters():
param_count += param.view(-1).size()[0]
print('total number of parameters of complete model: %d\n' % param_count)
# optimizer
optimizer = torch.optim.Adam(filter(lambda x: x.requires_grad is True, model.parameters()), lr=data.learning_rate)
# optimizer = torch.optim.Adam(model.parameters(), lr=data.learning_rate)
best_BLEU4 = -10
best_epoch = -1
for idx_ in range(int(data.epoch)):
idx = idx_ + 1
epoch_start = time.time()
temp_start = epoch_start
sample_loss = 0
total_loss = 0
model.train()
model.zero_grad()
random.shuffle(data.train_dataset)
batch_size = int(data.batch)
train_num = len(data.train_dataset)
total_batch = train_num // batch_size + 1
with tqdm(total=total_batch) as t:
for batch_id in range(total_batch):
t.set_description("EPOCH [{}/{}]".format(idx, data.epoch))
start = batch_id * batch_size
end = (batch_id + 1) * batch_size
if end > train_num:
end = train_num
instance = data.train_dataset[start:end]
if len(instance) == 0:
continue
src_tensor, tgt_tensor, skilltgt_tensor, skillnet_tensor, src_lengths, tgt_lengths, \
skill_tgt_lengths, skill_net_lengths = buildDataperBatch(instance, data.device, data.ifGPU)
loss = model.neg_log_likelihood_loss(
src_tensor, src_lengths, tgt_tensor, skilltgt_tensor, skillnet_tensor, skill_net_lengths)
t.set_postfix(loss=loss.item())
total_loss += loss.item()
loss.backward()
optimizer.step()
model.zero_grad()
t.update()
epoch_end = time.time()
epoch_cost = epoch_end - epoch_start
print("Epoch: %s training finished. Time: %.2fs, speed: %.2fst/s, total loss: %s" % (
idx, epoch_cost, train_num / epoch_cost, total_loss))
# evaluate
if idx >= 0:
BLUE_score_1, BLUE_score_2, BLUE_score_3, BLEU4 = evaluate(
data, model, data.result_dir + "/output_" + str(idx))
if BLEU4 > best_BLEU4:
best_epoch = idx
time_above = 0
print("best BLEU4 score\n BLEU1 %f, BLEU2 %f BLEU3 %f BLEU4 %f, epoch %d"
% (BLUE_score_1, BLUE_score_2, BLUE_score_3, BLEU4, best_epoch),
file=open(data.result_dir + "/BLEU_" + str(idx), "w"))
model_name = data.result_dir + "/modelFinal_{}.model".format(BLEU4)
torch.save(model.state_dict(), model_name)
best_BLEU4 = BLEU4
else:
time_above = time_above + 1
gc.collect()