-
Notifications
You must be signed in to change notification settings - Fork 7
/
adaptive_inference.py
172 lines (146 loc) · 6.65 KB
/
adaptive_inference.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
from __future__ import absolute_import
from __future__ import unicode_literals
from __future__ import print_function
from __future__ import division
import torch
import torch.nn as nn
import os
import math
def dynamic_evaluate(model, test_loader, val_loader, args):
tester = Tester(model, args)
if os.path.exists(os.path.join(args.save, 'logits_single.pth')):
val_pred, val_target, test_pred, test_target = \
torch.load(os.path.join(args.save, 'logits_single.pth'))
else:
val_pred, val_target = tester.calc_logit(val_loader)
test_pred, test_target = tester.calc_logit(test_loader)
torch.save((val_pred, val_target, test_pred, test_target),
os.path.join(args.save, 'logits_single.pth'))
flops = torch.load(os.path.join(args.save, 'flops.pth'))
with open(os.path.join(args.save, 'dynamic.txt'), 'w') as fout:
for p in range(1, 41):
print("*********************")
_p = torch.FloatTensor(1).fill_(p * 1.0 / 20)
# probs = [math.exp(math.log(p) * i) for i in in range(1, tester.args.nBlocks + 1)] # geometric distribution
probs = torch.exp(torch.log(_p) * torch.range(1, args.nBlocks))
probs /= probs.sum()
acc_val, _, T = tester.dynamic_eval_find_threshold(
val_pred, val_target, probs, flops)
acc_test, exp_flops = tester.dynamic_eval_with_threshold(
test_pred, test_target, flops, T)
print('valid acc: {:.3f}, test acc: {:.3f}, test flops: {:.2f}M'.format(acc_val, acc_test, exp_flops / 1e6))
fout.write('{}\t{}\n'.format(acc_test, exp_flops.item()))
class Tester(object):
def __init__(self, model, args=None):
self.args = args
self.model = model
self.softmax = nn.Softmax(dim=1).cuda()
def calc_logit(self, dataloader):
self.model.eval()
n_stage = self.args.nBlocks
logits = [[] for _ in range(n_stage)]
targets = []
for i, (input, target) in enumerate(dataloader):
targets.append(target)
with torch.no_grad():
input_var = torch.autograd.Variable(input)
output = self.model(input_var)
if not isinstance(output, list):
output = [output]
for b in range(n_stage):
_t = self.softmax(output[b])
# TODO check whether using the cpu type
# logits[b].append(_t.cpu())
logits[b].append(_t)
if i % self.args.print_freq == 0:
print('Generate Logit: [{0}/{1}]'.format(i, len(dataloader)))
# print(type(targets))
for b in range(n_stage):
logits[b] = torch.cat(logits[b], dim=0)
size = (n_stage, logits[0].size(0), logits[0].size(1))
ts_logits = torch.Tensor().resize_(size).zero_()
for b in range(n_stage):
ts_logits[b].copy_(logits[b])
targets = torch.cat(targets, dim=0)
ts_targets = torch.Tensor().resize_(size[1]).copy_(targets)
return ts_logits, ts_targets
def dynamic_eval_find_threshold(self, logits, targets, p, flops):
"""
logits: m * n * c
m: Stages
n: Samples
c: Classes
"""
n_stage, n_sample, c = logits.size()
max_preds, argmax_preds = logits.max(dim=2, keepdim=False)
_, sorted_idx = max_preds.sort(dim=1, descending=True)
filtered = torch.zeros(n_sample)
T = torch.Tensor(n_stage).fill_(1e8)
for k in range(n_stage - 1):
acc, count = 0.0, 0
out_n = math.floor(n_sample * p[k])
for i in range(n_sample):
ori_idx = sorted_idx[k][i]
if filtered[ori_idx] == 0:
count += 1
if count == out_n:
T[k] = max_preds[k][ori_idx]
break
filtered.add_(max_preds[k].ge(T[k]).type_as(filtered))
T[n_stage -1] = -1e8 # accept all of the samples at the last stage
acc_rec, exp = torch.zeros(n_stage), torch.zeros(n_stage)
acc, expected_flops = 0, 0
for i in range(n_sample):
gold_label = targets[i]
for k in range(n_stage):
if max_preds[k][i].item() >= T[k]: # force the sample to exit at k
if int(gold_label.item()) == int(argmax_preds[k][i].item()):
acc += 1
acc_rec[k] += 1
exp[k] += 1
break
acc_all = 0
for k in range(n_stage):
_t = 1.0 * exp[k] / n_sample
expected_flops += _t * flops[k]
acc_all += acc_rec[k]
# print('acc_all', acc_all * 1.0 / n_sample)
return acc * 100.0 / n_sample, expected_flops, T
def dynamic_eval_with_threshold(self, logits, targets, flops, T):
n_stage, n_sample, _ = logits.size()
max_preds, argmax_preds = logits.max(dim=2, keepdim=False) # take the max logits as confidence
acc_rec, exp = torch.zeros(n_stage), torch.zeros(n_stage)
acc, expected_flops = 0, 0
fout = open('index.txt', 'w')
for i in range(n_sample):
gold_label = targets[i]
for k in range(n_stage):
if max_preds[k][i].item() >= T[k]: # force to exit at k
_g = int(gold_label.item())
_pred = int(argmax_preds[k][i].item())
if _g == _pred:
# print(max_preds[k][i].item(), logits[k][i][_g].item())
acc += 1
acc_rec[k] += 1
"""
else:
fout.write('{}\t{}\t{}\t{}\t{:.3f}\n'.format(k, i, _pred, _g, max_preds[k][i].item()))
# print(max(logits[k][i]).item(), max_preds[k][i].item(), logits[k][i][_g].item())
"""
exp[k] += 1
break
acc_all, sample_all = 0, 0
for k in range(n_stage):
_t = exp[k] * 1.0 / n_sample
sample_all += exp[k]
expected_flops += _t * flops[k]
acc_all += acc_rec[k]
"""
for k in range(n_stage):
_t = exp[k] * 1.0 / n_sample
_acc = 1.0 * acc_rec[k] / exp[k]
print('stage[{}], exp: {:.3f}, acc: {:.3f}'.format(k, _t.item(), _acc.item()))
"""
# print('Test acc_all: {:.3f}, Case all: {}'.format(acc_all * 1.0 / n_sample, sample_all))
# print("Test acc_all", acc_all * 1.0 / n_sample)
return acc * 100.0 / n_sample, expected_flops