-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
176 lines (162 loc) · 7.3 KB
/
losses.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
import torch
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
def get_step_fn(std_value, train, conditioned=False, weighting=0):
"""Construct a one-step training/evaluation function.
Args:
std_value: (float/tuple) The standard deviation(s) for the noises.
train: (bool) The indication for training. It is set as True for training mode.
conditioned: (bool) The indication for the noise-conditioned model. It is set as
True if the model is conditioned on different standard deviations.
weighting: (int) The power of the balancing coefficient for the DSM loss. For example,
if wighting=2, the coefficient is 1/std^(2*wighting).
Returns:
step_fn: (func) A one-step training/evaluation function.
"""
def loss_fn(score_model, batch):
"""Compute the loss function.
Args:
score_model: (nn.Module) A parameterized score model.
batch: (tensor) A mini-batch of training data.
Returns:
loss: (float) The average loss value across the mini-batch.
"""
if conditioned:
# Calculate standard deviation
std_value_max, std_value_min = std_value
t = torch.rand(batch.shape[0], device=batch.device)
std_value_cond = (std_value_min * (std_value_max / std_value_min) ** t)
std = std_value_cond[:, None]
# Perturb the data
z = torch.randn_like(batch, device=batch.device)
perturbed_batch = batch + std * z
# Make predictions
score = score_model(perturbed_batch, std_value_cond)
else:
# Calculate standard deviation
std = torch.empty(batch.shape, device=batch.device).fill_(std_value)
# Perturb the data
z = torch.randn_like(batch, device=batch.device)
perturbed_batch = batch + std * z
# Make predictions
score = score_model(perturbed_batch)
# Calculate the losses
losses = torch.square(score * (std ** weighting) + z * (std ** (weighting-1)) )
loss = torch.mean(losses)
return loss
def step_fn(score_model, optimizer, batch):
"""Running one step of training or evaluation.
Args:
score_model: (nn.Module) A parameterized score model.
optimizer: (torch.optim) An optimizer function that can update score_model with '.step()' function.
batch: (tensor) A mini-batch of training data.
Returns:
loss: (float) The average loss value across the mini-batch.
"""
if train:
optimizer.zero_grad()
score_model.train()
loss = loss_fn(score_model, batch)
loss.backward()
optimizer.step()
else:
with torch.no_grad():
score_model.eval()
loss = loss_fn(score_model, batch)
return loss
return step_fn
def get_classifier_step_fn(std_value, train, loss_type='total', conditioned=False, weighting_dlsm=0, weighting_ce=0, coef=1.0, eps=1e-8):
"""Construct a one-step training/evaluation function.
Args:
std_value: (float/tuple) The standard deviation(s) for the noises.
loss_type: (str) The indication for the type of loss.
train: (bool) The indication for training. It is set as True for training mode.
conditioned: (bool) The indication for the noise-conditioned model. It is set as
True if the model is conditioned on different standard deviations.
weighting_dlsm: (int) The power of the balancing coefficient for the DLSM loss. For example,
if weighting_dlsm=2, the coefficient is 1/std^(2*2).
weighting_ce: (int) The power of the balancing coefficient for the CE loss. For example,
if weighting_ce=0, the coefficient is 1/std^(2*0).
coef: (float) The coefficient for balancing the DLSM and the CE losses.
eps: (float) An exetremely small value. It is used for preventing overflow.
Returns:
step_fn: (func) A one-step training/evaluation function.
"""
def loss_fn(classifier_model, score_model, batch, labels):
"""Compute the loss function.
Args:
score_model: (nn.Module) A parameterized score model.
classifier_model: (nn.Module) A parameterized classifier.
batch: (tensor) A mini-batch of training data.
labels: (tensor) A mini-batch of labels of the training data.
Returns:
loss: (float) The average loss value across the mini-batch.
"""
# Define softmax and ce functions
loss_ce_fn = torch.nn.CrossEntropyLoss(reduce=False)
sm = nn.Softmax(dim=1)
# Get standard deviation
if conditioned:
std_value_max, std_value_min = std_value
t = torch.rand(batch.shape[0], device=batch.device)
std_value_cond = (std_value_min * (std_value_max / std_value_min) ** t)
std = std_value_cond[:, None]
# Perturb the images
z = torch.randn_like(batch, device=batch.device)
perturbed_batch = batch + std * z
# Forward pass
with torch.no_grad():
score_model.eval()
score = score_model(perturbed_batch, std_value_cond)
perturbed_batch_var = Variable(perturbed_batch.clone(), requires_grad=True)
out = classifier_model(perturbed_batch_var, std_value_cond)
else:
std_value_cond = torch.empty(batch.shape[0], device=batch.device).fill_(std_value)
std = std_value_cond[:, None]
# Perturb the images
z = torch.randn_like(batch, device=batch.device)
perturbed_batch = batch + std * z
# Forward pass
with torch.no_grad():
score_model.eval()
score = score_model(perturbed_batch)
perturbed_batch_var = Variable(perturbed_batch.clone(), requires_grad=True)
out = classifier_model(perturbed_batch_var)
# Calculate the losses
if loss_type == 'total' or loss_type == 'dlsm':
# Calculate the dlsm loss
log_prob_class = torch.log(sm(out)+ eps)
label_mask = F.one_hot(labels, num_classes=2)
grads_prob_class, = torch.autograd.grad(log_prob_class, perturbed_batch_var,
grad_outputs=label_mask,
create_graph=True)
loss_dlsm = torch.mean(0.5 * torch.square(grads_prob_class * (std ** weighting_dlsm) + score * (std ** weighting_dlsm) + z * (std ** (weighting_dlsm-1)) ))
if loss_type == 'total' or loss_type == 'ce':
# Calculate the ce loss
loss_ce = torch.mean(loss_ce_fn(out, labels)*(std_value_cond ** (-2 * weighting_ce)))
loss = (loss_dlsm + coef * loss_ce) if loss_type == 'total' else (loss_dlsm if loss_type == 'dlsm' else loss_ce)
return loss
def step_fn(classifier_model, score_model, optimizer, batch, labels):
"""Running one step of training or evaluation.
Args:
score_model: (nn.Module) A parameterized score model.
classifier_model: (nn.Module) A parameterized classifier.
optimizer: (torch.optim) An optimizer function that can update score_model with '.step()' function.
batch: (tensor) A mini-batch of training data.
labels: (tensor) A mini-batch of labels of the training data.
Returns:
loss: (float) The average loss value across the mini-batch.
"""
if train:
optimizer.zero_grad()
classifier_model.train()
loss = loss_fn(classifier_model, score_model, batch, labels)
loss.backward()
optimizer.step()
else:
with torch.no_grad():
classifier_model.eval()
loss = loss_fn(classifier_model, score_model, batch, labels)
return loss
return step_fn