-
Notifications
You must be signed in to change notification settings - Fork 39
/
pseudo_labeling_util.py
159 lines (138 loc) · 6.79 KB
/
pseudo_labeling_util.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
import random
import time
import pickle
import numpy as np
import torch
import torch.nn.functional as F
from tqdm import tqdm
from .misc import AverageMeter, accuracy
from .utils import enable_dropout
def pseudo_labeling(args, data_loader, model, itr):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
end = time.time()
pseudo_idx = []
pseudo_target = []
pseudo_maxstd = []
gt_target = []
idx_list = []
gt_list = []
target_list = []
nl_mask = []
model.eval()
if not args.no_uncertainty:
f_pass = 10
enable_dropout(model)
else:
f_pass = 1
if not args.no_progress:
data_loader = tqdm(data_loader)
with torch.no_grad():
for batch_idx, (inputs, targets, indexs, _) in enumerate(data_loader):
data_time.update(time.time() - end)
inputs = inputs.to(args.device)
targets = targets.to(args.device)
out_prob = []
out_prob_nl = []
for _ in range(f_pass):
outputs = model(inputs)
out_prob.append(F.softmax(outputs, dim=1)) #for selecting positive pseudo-labels
out_prob_nl.append(F.softmax(outputs/args.temp_nl, dim=1)) #for selecting negative pseudo-labels
out_prob = torch.stack(out_prob)
out_prob_nl = torch.stack(out_prob_nl)
out_std = torch.std(out_prob, dim=0)
out_std_nl = torch.std(out_prob_nl, dim=0)
out_prob = torch.mean(out_prob, dim=0)
out_prob_nl = torch.mean(out_prob_nl, dim=0)
max_value, max_idx = torch.max(out_prob, dim=1)
max_std = out_std.gather(1, max_idx.view(-1,1))
out_std_nl = out_std_nl.cpu().numpy()
#selecting negative pseudo-labels
interm_nl_mask = ((out_std_nl < args.kappa_n) * (out_prob_nl.cpu().numpy() < args.tau_n)) *1
#manually setting the argmax value to zero
for enum, item in enumerate(max_idx.cpu().numpy()):
interm_nl_mask[enum, item] = 0
nl_mask.extend(interm_nl_mask)
idx_list.extend(indexs.numpy().tolist())
gt_list.extend(targets.cpu().numpy().tolist())
target_list.extend(max_idx.cpu().numpy().tolist())
#selecting positive pseudo-labels
if not args.no_uncertainty:
selected_idx = (max_value>=args.tau_p) * (max_std.squeeze(1) < args.kappa_p)
else:
selected_idx = max_value>=args.tau_p
pseudo_maxstd.extend(max_std.squeeze(1)[selected_idx].cpu().numpy().tolist())
pseudo_target.extend(max_idx[selected_idx].cpu().numpy().tolist())
pseudo_idx.extend(indexs[selected_idx].numpy().tolist())
gt_target.extend(targets[selected_idx].cpu().numpy().tolist())
loss = F.cross_entropy(outputs, targets.to(dtype=torch.long))
prec1, prec5 = accuracy(outputs[selected_idx], targets[selected_idx], topk=(1, 5))
losses.update(loss.item(), inputs.shape[0])
top1.update(prec1.item(), inputs.shape[0])
top5.update(prec5.item(), inputs.shape[0])
batch_time.update(time.time() - end)
end = time.time()
if not args.no_progress:
data_loader.set_description("Pseudo-Labeling Iter: {batch:4}/{iter:4}. Data: {data:.3f}s. Batch: {bt:.3f}s. Loss: {loss:.4f}. top1: {top1:.2f}. top5: {top5:.2f}. ".format(
batch=batch_idx + 1,
iter=len(data_loader),
data=data_time.avg,
bt=batch_time.avg,
loss=losses.avg,
top1=top1.avg,
top5=top5.avg,
))
if not args.no_progress:
data_loader.close()
pseudo_target = np.array(pseudo_target)
gt_target = np.array(gt_target)
pseudo_maxstd = np.array(pseudo_maxstd)
pseudo_idx = np.array(pseudo_idx)
#class balance the selected pseudo-labels
if itr < args.class_blnc-1:
min_count = 5000000 #arbitary large value
for class_idx in range(args.num_classes):
class_len = len(np.where(pseudo_target==class_idx)[0])
if class_len < min_count:
min_count = class_len
min_count = max(25, min_count) #this 25 is used to avoid degenarate cases when the minimum count for a certain class is very low
blnc_idx_list = []
for class_idx in range(args.num_classes):
current_class_idx = np.where(pseudo_target==class_idx)
if len(np.where(pseudo_target==class_idx)[0]) > 0:
current_class_maxstd = pseudo_maxstd[current_class_idx]
sorted_maxstd_idx = np.argsort(current_class_maxstd)
current_class_idx = current_class_idx[0][sorted_maxstd_idx[:min_count]] #select the samples with lowest uncertainty
blnc_idx_list.extend(current_class_idx)
blnc_idx_list = np.array(blnc_idx_list)
pseudo_target = pseudo_target[blnc_idx_list]
pseudo_idx = pseudo_idx[blnc_idx_list]
gt_target = gt_target[blnc_idx_list]
pseudo_labeling_acc = (pseudo_target == gt_target)*1
pseudo_labeling_acc = (sum(pseudo_labeling_acc)/len(pseudo_labeling_acc))*100
print(f'Pseudo-Labeling Accuracy (positive): {pseudo_labeling_acc}, Total Selected: {len(pseudo_idx)}')
pseudo_nl_mask = []
pseudo_nl_idx = []
nl_gt_list = []
for i in range(len(idx_list)):
if idx_list[i] not in pseudo_idx and sum(nl_mask[i]) > 0:
pseudo_nl_mask.append(nl_mask[i])
pseudo_nl_idx.append(idx_list[i])
nl_gt_list.append(gt_list[i])
nl_gt_list = np.array(nl_gt_list)
pseudo_nl_mask = np.array(pseudo_nl_mask)
one_hot_targets = np.eye(args.num_classes)[nl_gt_list]
one_hot_targets = one_hot_targets - 1
one_hot_targets = np.abs(one_hot_targets)
flat_pseudo_nl_mask = pseudo_nl_mask.reshape(1,-1)[0]
flat_one_hot_targets = one_hot_targets.reshape(1,-1)[0]
flat_one_hot_targets = flat_one_hot_targets[np.where(flat_pseudo_nl_mask == 1)]
flat_pseudo_nl_mask = flat_pseudo_nl_mask[np.where(flat_pseudo_nl_mask == 1)]
nl_accuracy = (flat_pseudo_nl_mask == flat_one_hot_targets)*1
nl_accuracy_final = (sum(nl_accuracy)/len(nl_accuracy))*100
print(f'Pseudo-Labeling Accuracy (negative): {nl_accuracy_final}, Total Selected: {len(nl_accuracy)}, Unique Samples: {len(pseudo_nl_mask)}')
pseudo_label_dict = {'pseudo_idx': pseudo_idx.tolist(), 'pseudo_target':pseudo_target.tolist(), 'nl_idx': pseudo_nl_idx, 'nl_mask': pseudo_nl_mask.tolist()}
return losses.avg, top1.avg, pseudo_labeling_acc, len(pseudo_idx), nl_accuracy_final, len(nl_accuracy), len(pseudo_nl_mask), pseudo_label_dict