-
Notifications
You must be signed in to change notification settings - Fork 22
/
rl_run.py
180 lines (147 loc) · 6.13 KB
/
rl_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
import tensorflow as tf
import os
from os import makedirs
from os.path import join, dirname
from time import strftime, gmtime
import numpy as np
from numpy import mean
from math import tanh
import json
from helpers import *
from cvae import CVAE
from classifier import EmojiClassifier
from emoji_reader import emoji_64
"""hyper-params"""
from params.full import *
num_epoch = 6
test_step = 20
emoji_num = 64
num_gpu = 1
"""directories"""
os.chdir("mojitalk_data")
output_dir = join("reinforced_cvae", strftime("%m-%d_%H-%M-%S", gmtime()))
vocab_f = "vocab.ori"
train_ori_f = "train.ori"
train_rep_f = "train.rep"
test_ori_f = "test.ori" # dev.ori
test_rep_f = "test.rep" # dev.rep
train_out_f = join(output_dir, "train.out")
test_out_f = join(output_dir, "test.out") # dev.out
makedirs(dirname(join(output_dir, "breakpoints/")), exist_ok=True)
log_f = open(join(output_dir, "log.txt"), "w", encoding="utf-8")
"""build vocabulary"""
word2index, index2word = build_vocab(vocab_f)
start_i, end_i = word2index['<s>'], word2index['</s>']
vocab_size = len(word2index)
emoji_b2s, emoji_s2b, emoji_sorted = build_emoji_index(vocab_f, emoji_64)
p = Printer(log_f, index2word)
"""build graphs and init params"""
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
graph0 = tf.Graph()
with graph0.as_default():
kl_ceiling = 0.48
seq2seq = CVAE(vocab_size, embed_size, num_unit, latent_dim, emoji_dim, batch_size,
kl_ceiling, 1, decoder_layer,
start_i, end_i, beam_width, maximum_iterations, max_gradient_norm, lr, dropout, num_gpu,
cell_type,
is_seq2seq=False)
sess0 = tf.Session(graph=graph0, config=config)
sess0.run(tf.global_variables_initializer())
seq2seq.set_sess(sess0)
train_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
saver0 = tf.train.Saver(var_list=train_vars, max_to_keep=25)
saver0.restore(sess0, "cvae/07-17_15-51-04/breakpoints/at_step_36500.ckpt")
graph0.finalize()
graph1 = tf.Graph()
with graph1.as_default():
classifier = EmojiClassifier(batch_size, vocab_size, emoji_num, embed_size, num_unit, num_gpu,
lr=0.001, dropout=0.2, cell_type=tf.nn.rnn_cell.GRUCell)
saver1 = tf.train.Saver()
sess1 = tf.Session(graph=graph1, config=config)
classifier.set_sess(sess1)
classifier.set_emoji_index(emoji_b2s)
saver1.restore(sess1, "classifier/07-16_14-33-58/breakpoints/best_test_loss.ckpt")
graph1.finalize()
"""build data"""
test_data = build_data(test_ori_f, test_rep_f, word2index)
test_batches = batch_generator(test_data, start_i, end_i, batch_size, permutate=False)
train_data = build_data(train_ori_f, train_rep_f, word2index)
global_step = best_step = 1
start_epoch = best_epoch = 1
total_step = (8 * len(train_data[0]) / batch_size)
(test_recon_loss, test_kl_loss, test_bow_loss,
perplexity, test_bleu_score, precisions, _) = seq2seq.infer_and_eval(test_batches)
p.put_example(seq2seq)
p.put_bleu(
test_recon_loss, test_kl_loss, test_bow_loss,
perplexity, test_bleu_score, precisions, "TEST")
lengths, ac, ac5 = seq2seq.policy_gen_eval(test_batches, classifier)
p.put_list([lengths, ac, ac5])
s = 1.
e = 1.
step = 500
for epoch in range(start_epoch, num_epoch + 1):
train_batches = batch_generator(train_data, start_i, end_i, batch_size, permutate=True)
l = [[] for i in range(5)]
for batch in train_batches:
emoji, ori, ori_len, rep, rep_len, _, _ = batch
gen, gen_len, normal, q_normal = seq2seq.get_generation(batch, get_normal=True)
with graph1.as_default():
gen_prob, gen_ac, gen_ac5 = classifier.get_prob(emoji, gen, gen_len)
tar_prob, tar_ac, tar_ac5 = classifier.get_prob(emoji, rep, rep_len)
gen_ac = gen_ac.astype(np.float32)
gen_ac5 = gen_ac5.astype(np.float32)
tar_ac = tar_ac.astype(np.float32)
tar_ac5 = tar_ac5.astype(np.float32)
reward = gen_prob - tar_prob
kl_w = get_kl_weight(global_step + 9500, total_step, 0.75)
pl_w = 1.05
policy_loss = seq2seq.policy_update(
reward, gen_ac, gen_ac5, normal, q_normal, batch,
min((e - s) / step * global_step + s, e),
pl_w,
kl_w)
reward *= (0.5 - 1) * gen_ac5 + 1
reward *= 1 - gen_ac
l[0].append(mean(gen_len))
l[1].append(mean(tar_ac) - mean(gen_ac))
l[2].append(mean(tar_ac5) - mean(gen_ac5))
l[3].append(np.sum(reward))
l[4].append(policy_loss)
if global_step % test_step == 0:
l = [mean(ll) for ll in l]
p.put_step(epoch, global_step)
p.put_list(l)
l = [[] for i in range(5)]
if global_step % (test_step * 20) == 10 * test_step:
(test_recon_loss, test_kl_loss, test_bow_loss,
perplexity, test_bleu_score, precisions, _) = seq2seq.infer_and_eval(test_batches)
p.put_example(seq2seq)
p.put_bleu(
test_recon_loss, test_kl_loss, test_bow_loss,
perplexity, test_bleu_score, precisions, "TEST")
if global_step % (test_step * 10) == 0:
p.put_step(epoch, global_step)
t_lengths, t_ac, t_ac5 = seq2seq.policy_gen_eval(test_batches, classifier)
p.put_list([t_lengths, t_ac, t_ac5])
if kl_w > 0.5 and t_ac5 > ac5:
path = join(output_dir, "breakpoints/at_step_%d.ckpt" % global_step)
save_path = saver0.save(sess0, path)
global_step += 1
"""RESTORE BEST MODEL"""
path = join(output_dir, "breakpoints/best_test_bleu.ckpt")
saver0.restore(sess0, path)
# TODO: eval train ac
"""GENERATE"""
# TRAIN SET
train_batches = batch_generator(
train_data, start_i, end_i, batch_size, permutate=False)
(train_recon_loss, train_kl_loss, train_bow_loss,
perplexity, train_bleu_score, precisions, generation_corpus) = seq2seq.infer_and_eval(train_batches)
write_out(train_out_f, generation_corpus, index2word)
p("BEST TRAIN BLEU: %.1f" % train_bleu_score)
# TEST SET
generation_corpus = seq2seq.infer_and_eval(test_batches)[-1]
write_out(test_out_f, generation_corpus, index2word)
log_f.close()