-
Notifications
You must be signed in to change notification settings - Fork 7
/
loss.py
167 lines (135 loc) · 5.11 KB
/
loss.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
'''
Loss functions.
'''
import copy
import torch
import torch.nn as nn
from torch.autograd import Variable
import numpy as np
import utils
class NLLLoss(nn.Module):
"""Self-Defined NLLLoss Function
Args:
weight: Tensor (num_class, )
"""
def __init__(self, weight):
super(NLLLoss, self).__init__()
self.weight = weight
def forward(self, prob, target):
"""
Args:
prob: (N, C)
target : (N, )
"""
N = target.size(0)
C = prob.size(1)
weight = Variable(self.weight).view((1, -1))
weight = weight.expand(N, C) # (N, C)
if prob.is_cuda:
weight = weight.cuda()
prob = weight * prob
one_hot = torch.zeros((N, C))
if prob.is_cuda:
one_hot = one_hot.cuda()
one_hot.scatter_(1, target.data.view((-1,1)), 1)
one_hot = one_hot.type(torch.ByteTensor)
one_hot = Variable(one_hot)
if prob.is_cuda:
one_hot = one_hot.cuda()
loss = torch.masked_select(prob, one_hot)
return -torch.sum(loss)
class GANLoss(nn.Module):
"""Reward-Refined NLLLoss Function for adversial training of Generator"""
def __init__(self):
super(GANLoss, self).__init__()
def forward_reinforce(self, prob, target, reward, cuda=False):
"""
Forward function used in the SeqGAN implementation.
Args:
prob: (N, C), torch Variable
target : (N, ), torch Variable
reward : (N, ), torch Variable
"""
N = target.size(0)
C = prob.size(1)
one_hot = torch.zeros((N, C))
if cuda:
one_hot = one_hot.cuda()
one_hot.scatter_(1, target.data.view((-1,1)), 1)
one_hot = one_hot.type(torch.ByteTensor)
one_hot = Variable(one_hot)
if cuda:
one_hot = one_hot.cuda()
loss = torch.masked_select(prob, one_hot)
loss = loss * reward
loss = -torch.sum(loss)
return loss
def forward_reward(self, i, samples, prob, rewards, BATCH_SIZE, g_sequence_len, VOCAB_SIZE, cuda=False):
"""
Returns what is used to get the gradient contribution of the i-th term of the batch.
"""
conditional_proba = Variable(torch.zeros(BATCH_SIZE, VOCAB_SIZE))
if cuda:
conditional_proba = conditional_proba.cuda()
for j in range(BATCH_SIZE):
conditional_proba[j, int(samples[j, i])] = 1
conditional_proba[j, :] = - (rewards[j]/BATCH_SIZE * conditional_proba[j, :])
return conditional_proba
def forward_reward_grads(self, samples, prob, rewards, g, BATCH_SIZE, g_sequence_len, VOCAB_SIZE, cuda=False):
"""
Returns a list of gradient contribution of every term in the batch
"""
conditional_proba = Variable(torch.zeros(BATCH_SIZE, g_sequence_len, VOCAB_SIZE))
batch_grads = []
if cuda:
conditional_proba = conditional_proba.cuda()
for j in range(BATCH_SIZE):
for i in range(g_sequence_len):
conditional_proba[j, i, int(samples[j, i])] = 1
conditional_proba[j, :, :] = - (rewards[j] * conditional_proba[j, :, :])
for j in range(BATCH_SIZE):
j_grads = []
# since we want to isolate each contribution, we have to zero the generator's gradients here.
g.zero_grad()
prob[j, :, :].backward(conditional_proba[j, :, :], retain_graph=True)
for p in g.parameters():
j_grads.append(p.grad.clone())
batch_grads.append(j_grads)
return batch_grads
class VarianceLoss(nn.Module):
"""Loss for the control variate annex network"""
def __init__(self):
super(VarianceLoss, self).__init__()
def forward(self, grad, cuda = False):
"""
Used to get the gradient of the variance.
"""
bs = len(grad)
ref = 0
for j in range(bs):
for i in range(len(grad[j])):
ref += torch.sum(grad[j][i]**2).item()
total_loss = np.array([ref/bs])
total_loss = Variable(torch.Tensor(total_loss), requires_grad=True)
if cuda:
total_loss = total_loss.cuda()
return total_loss
def forward_variance(self, grad, cuda=False):
"""
Used to get the variance of one single parameter.
In this case, we take look at the last layer, then take the variance of the first parameter of this last layer in main.py
"""
bs = len(grad)
n_layers = len(grad[0])
square_term = torch.zeros((grad[0][n_layers-1].size()))
normal_term = torch.zeros((grad[0][n_layers-1].size()))
if cuda:
square_term = square_term.cuda()
normal_term = normal_term.cuda()
for j in range(bs):
square_term = torch.add(square_term, grad[j][n_layers-1]**2)
normal_term = torch.add(normal_term, grad[j][n_layers-1])
square_term /= bs
normal_term /= bs
normal_term = normal_term ** 2
return square_term - normal_term