-
Notifications
You must be signed in to change notification settings - Fork 4
/
loss_functions.py
194 lines (155 loc) · 7.01 KB
/
loss_functions.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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
"""
Edanur Demir
Loss functions used in EENet training
"""
import sys
import torch
import torch.nn.functional as F
def loss(args, exit_tag, pred, target, conf, cost):
"""loss function
Arguments are
* args: command line arguments entered by user.
* pred: prediction result of each exit point.
* target: target prediction values.
* conf: confidence value of each exit point.
* cost: cost rate of the each exit point.
This function switches between the loss functions.
"""
if args.loss_func == 'v0':
return loss_v0(args, pred, target, conf, cost)
if args.loss_func == 'v1':
return loss_v1(args, pred, target, conf, cost)
if args.loss_func == 'v2':
return loss_v2(args, pred, target, conf, cost)
if args.loss_func == 'v3':
return loss_v3(args, pred, target, conf, cost)
if args.loss_func == 'v4':
return loss_v4(args, exit_tag, pred, target, conf, cost)
def loss_v0(args, pred, target, conf, cost):
"""loss version 0
Arguments are
* args: command line arguments entered by user.
* pred: prediction result of each exit point.
* target: target prediction values.
* conf: confidence value of each exit point.
* cost: cost rate of the each exit point.
This loss function is the cumulative loss of all exit points.
It is used in the first stage of two-stage training.
"""
pred_loss = 0
cost_loss = 0
for i in range(args.num_ee + 1):
pred_loss += F.nll_loss(pred[i].log(), target)
cum_loss = pred_loss
return cum_loss, pred_loss, cost_loss
def loss_v1(args, pred, target, conf, cost):
"""loss version 1
Arguments are
* args: command line arguments entered by user.
* pred: prediction result of each exit point.
* target: target prediction values.
* conf: confidence value of each exit point.
* cost: cost rate of the each exit point.
This loss function is the fusion loss of the cross_entropy loss and cost loss.
These loss parts are calculated in a recursive way as following:
Prediction'_i = confidence_i * prediction_i + (1 - confidence_i) * Prediction'_(i+1)
Cost'_i = confidence_i * cost_i + (1 - confidence_i) * Cost'_(i+1)
"""
cum_pred = pred[args.num_ee]
cum_cost = cost[args.num_ee]
for i in range(args.num_ee-1, -1, -1):
cum_pred = conf[i] * pred[i] + (1-conf[i]) * cum_pred
cum_cost = conf[i] * cost[i] + (1-conf[i]) * cum_cost
pred_loss = F.nll_loss(cum_pred.log(), target)
cost_loss = cum_cost.mean()
cum_loss = pred_loss + args.lambda_coef * cost_loss
return cum_loss, pred_loss, cost_loss
def loss_v2(args, pred, target, conf, cost):
"""loss version 2
Arguments are
* args: command line arguments entered by user.
* pred: prediction result of each exit point.
* target: target prediction values.
* conf: confidence value of each exit point.
* cost: cost rate of the each exit point.
This loss function is the cumulative loss of loss_v1 by recursively.
It aims to provide a more fair training.
"""
cum_pred = [None] * args.num_ee + [pred[args.num_ee]]
cum_cost = [None] * args.num_ee + [cost[args.num_ee]]
pred_loss = F.nll_loss(cum_pred[-1].log(), target)
cum_loss = pred_loss + args.lambda_coef * cum_cost[-1].mean()
for i in range(args.num_ee-1, -1, -1):
cum_pred[i] = conf[i] * pred[i] + (1-conf[i]) * cum_pred[i+1]
cum_cost[i] = conf[i] * cost[i] + (1-conf[i]) * cum_cost[i+1]
pred_loss = F.nll_loss(cum_pred[i].log(), target)
cost_loss = cum_cost[i].mean()
cum_loss += pred_loss + args.lambda_coef * cost_loss
return cum_loss, 0, 0
def loss_v3(args, pred, target, conf, cost):
"""loss version 3
Arguments are
* args: command line arguments entered by user.
* pred: prediction result of each exit point.
* target: target prediction values.
* conf: confidence value of each exit point.
* cost: cost rate of the each exit point.
This loss function uses the normalized confidence values.
"""
conf_sum = 0
for i in range(len(conf)):
conf_sum += conf[i].mean()
norm_conf = [conf[i].mean() / conf_sum for i in range(len(conf))]
cum_loss = 0
for i in range(args.num_ee + 1):
pred_loss = F.nll_loss(pred[i].log(), target)
cost_loss = cost[i].mean()
cum_loss += norm_conf[i] * (pred_loss + args.lambda_coef * cost_loss)
return cum_loss, 0, 0
def loss_v4(args, exit_tag, pred, target, conf, cost):
"""loss version 4
Arguments are
* args: command line arguments entered by user.
* exit_tag: exit tag of examples in the batch.
* pred: prediction result of each exit point.
* target: target prediction values.
* conf: confidence value of each exit point.
* cost: cost rate of the each exit point.
This loss function uses the exit tags of examples pre-assigned by the model.
"""
cum_pred = pred[args.num_ee]
cum_cost = cost[args.num_ee]
conf_loss = 0
for i in range(args.num_ee + 1):
exiting_examples = (exit_tag == i).to(args.device, dtype=torch.float)
not_exiting_examples = (exit_tag != i).to(args.device, dtype=torch.float)
cum_pred = exiting_examples * pred[i] + not_exiting_examples * cum_pred
cum_cost = exiting_examples * cost[i] + not_exiting_examples * cum_cost
exiting_rate = exiting_examples.sum().item() / len(exit_tag)
not_exiting_rate = not_exiting_examples.sum() / len(exit_tag)
conf_weights = exiting_examples * not_exiting_rate + not_exiting_examples * exiting_rate
conf_loss += F.binary_cross_entropy(conf[i], exiting_examples, conf_weights)
pred_loss = F.nll_loss(cum_pred.log(), target)
cost_loss = cum_cost.mean()
cum_loss = pred_loss + args.lambda_coef * cost_loss + conf_loss
return cum_loss, pred_loss, cost_loss
def update_exit_tags(args, batch_size, pred, target, cost):
"""loss version 4
Arguments are
* args: command line arguments entered by user.
* batch_size: current size of the batch.
* pred: prediction result of each exit point.
* target: target prediction values.
* cost: cost rate of the each exit point.
This function updates and returns the exit tags.
"""
cum_loss = (torch.ones(batch_size) * sys.maxsize).to(args.device)
exit_tag = (torch.ones(batch_size) * args.num_ee).to(args.device, dtype=torch.int)
for exit in range(args.num_ee + 1):
loss = F.nll_loss(pred[exit].log(), target, reduction='none') \
+ args.lambda_coef * cost[exit]
smaller_values = (loss < cum_loss).to(args.device, dtype=torch.float)
greater_values = (loss >= cum_loss).to(args.device, dtype=torch.float)
cum_loss = loss * smaller_values + cum_loss * greater_values
exit_tag = exit * smaller_values.int() + exit_tag * greater_values.int()
return exit_tag.reshape(-1, 1)