-
Notifications
You must be signed in to change notification settings - Fork 5
/
losses.py
158 lines (129 loc) · 6.41 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
import torch
from torch.nn import functional as F
class DistillationLoss(torch.nn.Module):
"""
This module wraps a standard criterion and adds an extra knowledge distillation loss by
taking a teacher model prediction and using it as additional supervision.
"""
def __init__(self, base_criterion: torch.nn.Module, teacher_model: torch.nn.Module,
distillation_type: str, alpha: float, tau: float, kl_avg: str = "batchmean"):
super().__init__()
self.base_criterion = base_criterion
self.teacher_model = teacher_model
assert distillation_type in ['none', 'soft', 'hard']
self.distillation_type = distillation_type
self.alpha = alpha
self.tau = tau
self.kl_avg = kl_avg
def distillation_func(self, inputs, outputs_kd):
if outputs_kd is None:
raise ValueError("When knowledge distillation is enabled, the model is "
"expected to return a Tuple[Tensor, Tensor] with the output of the "
"class_token and the dist_token")
# don't backprop throught the teacher
with torch.no_grad():
teacher_outputs = self.teacher_model(inputs)
if self.distillation_type == 'soft':
T = self.tau
# taken from https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
# with slight modifications
distillation_loss = F.kl_div(
F.log_softmax(outputs_kd / T, dim=1),
#We provide the teacher's targets in log probability because we use log_target=True
#(as recommended in pytorch https://github.com/pytorch/pytorch/blob/9324181d0ac7b4f7949a574dbc3e8be30abe7041/torch/nn/functional.py#L2719)
#but it is possible to give just the probabilities and set log_target=False. In our experiments we tried both.
F.log_softmax(teacher_outputs / T, dim=1),
reduction=self.kl_avg,
log_target=True
) * (T * T)
elif self.distillation_type == 'hard':
distillation_loss = F.cross_entropy(outputs_kd, teacher_outputs.argmax(dim=1))
return distillation_loss
def forward(self, inputs, outputs, labels, model=None):
"""
Args:
inputs: The original inputs that are feed to the teacher model
outputs: the outputs of the model to be trained. It is expected to be
either a Tensor, or a Tuple[Tensor, Tensor], with the original output
in the first position and the distillation predictions as the second output
labels: the labels for the base criterion
"""
outputs_kd = None
if not isinstance(outputs, torch.Tensor):
# assume that the model outputs a tuple of [outputs, outputs_kd]
outputs, outputs_kd = outputs
base_loss = self.base_criterion(outputs, labels)
if self.distillation_type == 'none':
return base_loss
distillation_loss = self.distillation_func(inputs, outputs_kd)
loss = base_loss * (1 - self.alpha) + distillation_loss * self.alpha
return loss
class DynamicViTDistillationLoss(torch.nn.Module):
"""
This module wraps a standard criterion and adds an extra knowledge distillation loss by
taking a teacher model prediction and using it as additional supervision.
"""
def __init__(self, base_criterion: torch.nn.Module, teacher_model: torch.nn.Module,
ratio_weight: float, cls_distill_weight: float, token_distill_weight: float,
cls_weight: float, mse_token: bool, kl_avg: str = "batchmean"):
super().__init__()
self.teacher_model = teacher_model
self.base_criterion = base_criterion
self.cls_weight = cls_weight
self.cls_distill_weight = cls_distill_weight
self.token_distill_weight = token_distill_weight
self.ratio_weight = ratio_weight
self.mse_token = mse_token
self.kl_avg = kl_avg
def forward(self, inputs, outputs, labels, model):
"""
Args:
inputs: The original inputs that are feed to the teacher model
outputs: the outputs of the model to be trained. It is expected to be
either a Tensor, or a Tuple[Tensor, Tensor], with the original output
in the first position and the distillation predictions as the second output
labels: the labels for the base criterion
"""
pred, token_pred, mask, out_pred_score = outputs
base_loss = self.base_criterion(pred, labels)
loss = base_loss * self.cls_weight
pred_loss = 0.0
try:
keep_rate = model.module.token_ratio
except AttributeError:
keep_rate = model.token_ratio
for i, score in enumerate(out_pred_score):
pred_loss = pred_loss + ((score.mean(1) - keep_rate[i]) ** 2).mean()
pred_loss /= len(out_pred_score)
loss += pred_loss * self.ratio_weight
if self.teacher_model is not None:
with torch.no_grad():
cls_t, token_t = self.teacher_model(inputs)
cls_kl_loss = F.kl_div(
F.log_softmax(pred, dim=-1),
F.log_softmax(cls_t, dim=-1),
reduction=self.kl_avg,
log_target=True
)
loss += self.cls_distill_weight * cls_kl_loss
B, N, C = token_pred.shape
assert mask.numel() == B * N
bool_mask = mask.reshape(B*N) > 0.5
token_pred = token_pred.reshape(B*N, C)
token_t = token_t.reshape(B*N, C)
if mask.sum() < 0.1:
token_kl_loss = token_pred.new(1,).fill_(0.0)
else:
token_t = token_t[bool_mask]
token_pred = token_pred[bool_mask]
if self.mse_token:
token_kl_loss = torch.pow(token_pred - token_t, 2).mean()
else:
token_kl_loss = F.kl_div(
F.log_softmax(token_pred, dim=-1),
F.log_softmax(token_t, dim=-1),
reduction=self.kl_avg,
log_target=True
)
loss += self.token_distill_weight * token_kl_loss
return loss