-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathadversarial.py
181 lines (141 loc) · 6.68 KB
/
adversarial.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
# coding:utf-8
import os
import time
import pickle
import argparse
import logging
import tqdm_logging
from utils import SYM_PAD
from data import batcher, load_vocab, padding_inputs, sentence2id
from encoder import EncoderRNN
from generator import Generator
from discriminator import Discriminator
from toy import save_model, reload_model
import torch
import torch.nn as nn
def adversarial():
# user the root logger
logger = logging.getLogger("lan2720")
argparser = argparse.ArgumentParser(add_help=False)
argparser.add_argument('--load_path', '-p', type=str, required=True)
# TODO: load best
argparser.add_argument('--load_epoch', '-e', type=int, required=True)
argparser.add_argument('--filter_num', type=int, required=True)
argparser.add_argument('--filter_sizes', type=str, required=True)
argparser.add_argument('--training_ratio', type=int, default=2)
argparser.add_argument('--g_learning_rate', '-glr', type=float, default=0.001)
argparser.add_argument('--d_learning_rate', '-dlr', type=float, default=0.001)
argparser.add_argument('--batch_size', '-b', type=int, default=168)
# new arguments used in adversarial
new_args = argparser.parse_args()
# load default arguments
default_arg_file = os.path.join(new_args.load_path, 'args.pkl')
if not os.path.exists(default_arg_file):
raise RuntimeError('No default argument file in %s' % new_args.load_path)
else:
with open(default_arg_file, 'rb') as f:
args = pickle.load(f)
args.mode = 'adversarial'
#args.d_learning_rate = 0.0001
args.print_every = 1
args.g_learning_rate = new_args.g_learning_rate
args.d_learning_rate = new_args.d_learning_rate
args.batch_size = new_args.batch_size
# add new arguments
args.load_path = new_args.load_path
args.load_epoch = new_args.load_epoch
args.filter_num = new_args.filter_num
args.filter_sizes = new_args.filter_sizes
args.training_ratio = new_args.training_ratio
# set up the output directory
exp_dirname = os.path.join(args.exp_dir, args.mode, time.strftime("%Y-%m-%d-%H-%M-%S"))
os.makedirs(exp_dirname)
# set up the logger
tqdm_logging.config(logger, os.path.join(exp_dirname, 'adversarial.log'),
mode='w', silent=False, debug=True)
# load vocabulary
vocab, rev_vocab = load_vocab(args.vocab_file, max_vocab=args.max_vocab_size)
vocab_size = len(vocab)
word_embeddings = nn.Embedding(vocab_size, args.emb_dim, padding_idx=SYM_PAD)
E = EncoderRNN(vocab_size, args.emb_dim, args.hidden_dim, args.n_layers, args.dropout_rate, bidirectional=True, variable_lengths=True)
G = Generator(vocab_size, args.response_max_len, args.emb_dim, 2*args.hidden_dim, args.n_layers, dropout_p=args.dropout_rate)
D = Discriminator(args.emb_dim, args.filter_num, eval(args.filter_sizes))
if args.use_cuda:
word_embeddings.cuda()
E.cuda()
G.cuda()
D.cuda()
# define optimizer
opt_G = torch.optim.Adam(G.rnn.parameters(), lr=args.g_learning_rate)
opt_D = torch.optim.Adam(D.parameters(), lr=args.d_learning_rate)
logger.info('----------------------------------')
logger.info('Adversarial a neural conversation model')
logger.info('----------------------------------')
logger.info('Args:')
logger.info(str(args))
logger.info('Vocabulary from ' + args.vocab_file)
logger.info('vocabulary size: %d' % vocab_size)
logger.info('Loading text data from ' + args.train_query_file + ' and ' + args.train_response_file)
reload_model(args.load_path, args.load_epoch, word_embeddings, E, G)
# start_epoch = args.resume_epoch + 1
#else:
# start_epoch = 0
# dump args
with open(os.path.join(exp_dirname, 'args.pkl'), 'wb') as f:
pickle.dump(args, f)
# TODO: num_epoch is old one
for e in range(args.num_epoch):
train_data_generator = batcher(args.batch_size, args.train_query_file, args.train_response_file)
logger.info("Epoch: %d/%d" % (e, args.num_epoch))
step = 0
cur_time = time.time()
while True:
try:
post_sentences, response_sentences = train_data_generator.next()
except StopIteration:
# save model
save_model(exp_dirname, e, word_embeddings, E, G, D)
## evaluation
#eval(args.valid_query_file, args.valid_response_file, args.batch_size,
# word_embeddings, E, G, loss_func, args.use_cuda, vocab, args.response_max_len)
break
# prepare data
post_ids = [sentence2id(sent, vocab) for sent in post_sentences]
response_ids = [sentence2id(sent, vocab) for sent in response_sentences]
posts_var, posts_length = padding_inputs(post_ids, None)
responses_var, responses_length = padding_inputs(response_ids, args.response_max_len)
# sort by post length
posts_length, perms_idx = posts_length.sort(0, descending=True)
posts_var = posts_var[perms_idx]
responses_var = responses_var[perms_idx]
responses_length = responses_length[perms_idx]
if args.use_cuda:
posts_var = posts_var.cuda()
responses_var = responses_var.cuda()
embedded_post = word_embeddings(posts_var)
real_responses = word_embeddings(responses_var)
# forward
_, dec_init_state = E(embedded_post, input_lengths=posts_length.numpy())
fake_responses = G(dec_init_state, word_embeddings) # [B, T, emb_size]
prob_real = D(embedded_post, real_responses)
prob_fake = D(embedded_post, fake_responses)
# loss
D_loss = - torch.mean(torch.log(prob_real) + torch.log(1. - prob_fake))
G_loss = torch.mean(torch.log(1. - prob_fake))
if step % args.training_ratio == 0:
opt_D.zero_grad()
D_loss.backward(retain_graph=True)
opt_D.step()
opt_G.zero_grad()
G_loss.backward()
opt_G.step()
if step % args.print_every == 0:
logger.info('Step %5d: D accuracy=%.2f (0.5 for D to converge) D score=%.2f (-1.38 for G to converge) (%.1f iters/sec)' % (
step,
prob_real.cpu().data.numpy().mean(),
-D_loss.cpu().data.numpy()[0],
args.print_every/(time.time()-cur_time)))
cur_time = time.time()
step = step + 1
if __name__ == '__main__':
adversarial()