-
Notifications
You must be signed in to change notification settings - Fork 35
/
model.py
162 lines (145 loc) · 6.89 KB
/
model.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
import random
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.distributions import Categorical
class Policy(nn.Module):
def __init__(self, args, n_class, base_model, in_dim):
super(Policy, self).__init__()
self.args = args
self.L = 0.02
self.baseline_reward = 0
self.entropy_l = []
self.beta = args.beta
self.beta_decay_rate = .9
self.n_update = 0
self.class_embed = nn.Embedding(n_class, args.class_embed_size)
self.class_embed_bias = nn.Embedding(n_class, 1)
stdv = 1. / np.sqrt(self.class_embed.weight.size(1))
self.class_embed.weight.data.uniform_(-stdv, stdv)
self.class_embed_bias.weight.data.uniform_(-stdv, stdv)
self.saved_log_probs = []
self.rewards = []
self.rewards_greedy = []
self.doc_vec = None
self.base_model = base_model
self.criterion = torch.nn.BCEWithLogitsLoss()
self.sl_loss = 0
if self.args.use_history:
self.state_hist = None
self.output_hist = None
self.hist_gru = nn.GRU(args.class_embed_size, args.class_embed_size, bidirectional=True)
if self.args.use_cur_class_embed:
in_dim += self.args.class_embed_size
if self.args.use_history:
in_dim += args.hist_embed_size * 2
if self.args.use_l2:
self.l1 = nn.Linear(in_dim, args.l1_size)
self.l2 = nn.Linear(args.l1_size, args.class_embed_size)
elif self.args.use_l1:
self.l1 = nn.Linear(in_dim, args.class_embed_size)
def update_baseline(self, target):
# a moving average baseline, not used anymore
self.baseline_reward = self.L * target + (1 - self.L) * self.baseline_reward
def finish_episode(self):
self.sl_loss = 0
self.n_update += 1
if self.n_update % self.args.update_beta_every == 0:
self.beta *= self.beta_decay_rate
self.entropy_l = []
del self.rewards[:]
del self.saved_log_probs[:]
del self.rewards_greedy[:]
def forward(self, cur_class_batch, next_classes_batch):
cur_class_embed = self.class_embed(cur_class_batch) # (batch, 50)
next_classes_embed = self.class_embed(next_classes_batch) # (batch, max_choices, 50)
nb = self.class_embed_bias(next_classes_batch).squeeze(-1)
states_embed = self.doc_vec
if self.args.use_cur_class_embed:
states_embed = torch.cat((states_embed, cur_class_embed), 1)
if self.args.use_history:
states_embed = torch.cat((states_embed, self.output_hist.squeeze()), 1)
if not self.args.use_l1:
return torch.bmm(next_classes_embed, states_embed.unsqueeze(-1)).squeeze(-1) + nb
if self.args.use_l2:
h1 = F.relu(self.l1(states_embed))
h2 = F.relu(self.l2(h1))
else:
h2 = F.relu(self.l1(states_embed))
h2 = h2.unsqueeze(-1) # (batch, 50, 1)
probs = torch.bmm(next_classes_embed, h2).squeeze(-1) + nb
if self.args.use_history:
self.output_hist, self.state_hist = self.hist_gru(cur_class_embed.unsqueeze(0), self.state_hist)
return probs
def duplicate_doc_vec(self, indices):
assert self.doc_vec is not None
assert len(indices) > 0
self.doc_vec = self.doc_vec[indices]
def duplicate_reward(self, indices):
assert len(indices) > 0
self.saved_log_probs[-1] = [[probs[i] for i in indices] for probs in self.saved_log_probs[-1]]
self.rewards[-1] = [[R[i] for i in indices] for R in self.rewards[-1]]
def generate_doc_vec(self, mini_batch):
self.doc_vec = self.base_model(mini_batch)
def generate_logits(self, mini_batch, cur_class_batch, next_classes_batch):
if self.doc_vec is None:
self.generate_doc_vec(mini_batch)
if self.args.gpu:
cur_class_batch = Variable(torch.from_numpy(cur_class_batch)).cuda()
next_classes_batch = Variable(torch.from_numpy(next_classes_batch)).cuda()
else:
cur_class_batch = Variable(torch.from_numpy(cur_class_batch))
next_classes_batch = Variable(torch.from_numpy(next_classes_batch))
logits = self(cur_class_batch, next_classes_batch)
# mask padding relations
logits = (next_classes_batch == 0).float() * -99999 + (next_classes_batch != 0).float() * logits
return logits
def step_sl(self, mini_batch, cur_class_batch, next_classes_batch, next_classes_batch_true, sigmoid=True):
logits = self.generate_logits(mini_batch, cur_class_batch, next_classes_batch)
if not sigmoid:
return logits
if next_classes_batch_true is not None:
if self.args.gpu:
y_true = Variable(torch.from_numpy(next_classes_batch_true)).cuda().float()
else:
y_true = Variable(torch.from_numpy(next_classes_batch_true)).float()
self.sl_loss += self.criterion(logits, y_true)
return F.sigmoid(logits)
def step(self, mini_batch, cur_class_batch, next_classes_batch, test=False, flat_probs=None):
logits = self.generate_logits(mini_batch, cur_class_batch, next_classes_batch)
if self.args.softmax:
probs = F.softmax(logits, dim=-1)
else:
probs = F.sigmoid(logits)
if not test:
# + epsilon to avoid log(0)
self.entropy_l.append(torch.mean(torch.log(probs + 1e-32) * probs))
next_classes_batch = Variable(torch.from_numpy(next_classes_batch)).cuda()
probs = probs + (next_classes_batch != 0).float() * 1e-16
m = Categorical(probs)
if test or self.args.sample_mode == 'choose_max':
action = torch.max(probs, 1)[1]
elif self.args.sample_mode == 'random':
if random.random() < 1.2:
if self.args.gpu:
action = Variable(torch.zeros(probs.size()[0]).long().random_(0, probs.size()[1])).cuda()
else:
action = Variable(torch.zeros(probs.size()[0]).long().random_(0, probs.size()[1]))
else:
action = m.sample()
else:
action = m.sample()
return action, m
# not used anymore
def init_hist(self, tokens_size):
if self.args.gpu:
self.state_hist = self.init_hidden(tokens_size, self.args.hist_embed_size).cuda()
first_input = Variable(
torch.from_numpy(np.zeros((1, tokens_size, self.args.class_embed_size)))).cuda().float()
else:
self.state_hist = self.init_hidden(tokens_size, self.args.hist_embed_size)
first_input = Variable(
torch.from_numpy((np.zeros(1, tokens_size, self.args.class_embed_size)))).float()
self.output_hist, self.state_hist = self.hist_gru(first_input, self.state_hist)