-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathreinforce.py
131 lines (96 loc) · 4.16 KB
/
reinforce.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
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
import torch
from torch.autograd import Variable
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
from torch.distributions import Categorical
import policy_config
conf = policy_config.config()
USE_CUDA = torch.cuda.is_available()
FloatTensor = torch.cuda.FloatTensor if USE_CUDA else torch.FloatTensor
LongTensor = torch.cuda.LongTensor if USE_CUDA else torch.LongTensor
ByteTensor = torch.cuda.ByteTensor if USE_CUDA else torch.ByteTensor
def reward_length(x):
return 16*x**2*(1-x)**2
def reward_grammar(word_seq_id, dep_seq_id, pos_seq_id, language_model, pred_actions, mask, indexs):
new_word_seq_id = np.array(word_seq_id)[indexs].tolist()
new_dep_seq_id = np.array(dep_seq_id)[indexs].tolist()
new_pos_seq_id = np.array(pos_seq_id)[indexs].tolist()
word2id = language_model.word2id
dep2id = language_model.dep2id
pos2id = language_model.pos2id
new_word_seq_id1 = []
new_dep_seq_id1 = []
new_pos_seq_id1 = []
for i in range(len(pred_actions)):
temp_word = [word2id['BOS']]
temp_dep = [dep2id['BOS']]
temp_pos = [pos2id['BOS']]
for j in range(sum(mask[i])):
if pred_actions[i][j]==1:
temp_word.append(new_word_seq_id[i][j])
temp_dep.append(new_dep_seq_id[i][j])
temp_pos.append(new_pos_seq_id[i][j])
temp_word.append(word2id['EOS'])
temp_dep.append(dep2id['EOS'])
temp_pos.append(pos2id['EOS'])
new_word_seq_id1.append(temp_word)
new_dep_seq_id1.append(temp_dep)
new_pos_seq_id1.append(temp_pos)
new_target_seq_id1=[]
for sent in new_word_seq_id1:
new_target_seq_id1.append(sent[1:-1])
h0 = language_model.init_hidden(conf.batch_size)
_, probs, _, target_padded_ids, indexs, mask = \
language_model(new_word_seq_id1, new_dep_seq_id1, new_pos_seq_id1, new_target_seq_id1, h0, False)
ori_index = np.argsort(indexs).tolist()
probs=probs[ori_index].cpu().data.numpy()
mask=mask[ori_index].cpu().data.numpy()
target_padded_ids=target_padded_ids[ori_index].cpu().data.numpy()
print
ppl=[]
for i in range(len(mask)):
l=int(sum(mask[i]))
temp=[]
for probability, id_ in zip(probs[i][:l], target_padded_ids[i][:l]):
print probability[id_]
temp.append(np.log(probability[id_]))
ppl.append(np.exp(-sum(temp)/float(len(temp))))
rewards = FloatTensor(ppl)
rewards = (rewards - rewards.mean()) / (rewards.std() + np.finfo(np.float32).eps)
return rewards
def select_action(word_seq_id, dep_seq_id, pos_seq_id, policy, language_model, optimizer):
h0 = policy.init_hidden(conf.batch_size)
_, probs, indexs, mask, pred = policy(word_seq_id, dep_seq_id, pos_seq_id, h0, True)
pred_actions=[]
for prob, index, mask_i in zip(probs, indexs, mask):
length = sum(mask_i).astype('int64')
m = Categorical(prob[:length])
actions = m.sample()
pred_actions.append(actions.data.numpy().tolist()+[-1]*(mask.shape[1]-length))
policy.saved_log_probs.append(m.log_prob(actions))
policy.pred_actions=pred_actions
policy.length_r = reward_length(model.length_r)
policy.grammar_r =reward_grammar(word_seq_id, dep_seq_id, pos_seq_id, language_model,
pred_actions, mask, indexs)
batch_loss=0
for i in range(mask.shape[0]):
batch_loss+=finish_episode(policy, i)#/float(length)
batch_loss = batch_loss/conf.batch_size
optimizer.zero_grad()
batch_loss.backward()
optimizer.step()
policy.saved_log_probs = []
policy.length_r = []
policy.grammar_r = []
return batch_loss
def finish_episode(model, index):
model_loss = []
for log_prob in model.saved_log_probs[index]:
model_loss.append(-log_prob * (model.grammar_r[index]+model.length_r[index]))
model_loss = torch.cat(model_loss).sum()
return model_loss