-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmisc.py
181 lines (149 loc) · 7.09 KB
/
misc.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
import torch
import torch.nn as nn
import numpy as np
from data_loader import get_loader,CelebA
from torchvision.models import inception_v3
import logger
import time,datetime
import random
import torch.nn.functional as F
import sys,os
from torchvision import transforms as T
class InceptionNet():
def __init__(self,config):
self.image_size=299 #Inception net condition
self.lr=0.0001
self.log_step=100
self.selected_attrs=config.selected_attrs
self.device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.buildIncNet()
self.save_incDir=config.inc_net_dir
self.pretrained_incNet=config.pretrained_incNet
self.dataset=config.dataset
self.test_dataset=get_loader(config.celeba_image_dir, config.attr_path,
config.selected_attrs,image_size=self.image_size,num_workers=config.num_workers,
dataset=config.dataset,mode='test')
def buildIncNet(self):
self.inc_net=inception_v3(pretrained=False, num_classes=len(self.selected_attrs),aux_logits=False)
self.opt=torch.optim.Adam(self.inc_net.parameters(),self.lr,[0.5,0.999])
self.inc_net.to(self.device)
def load_pretrained(self):
if self.pretrained_incNet is not None:
self.inc_net.load_state_dict(torch.load(self.pretrained_incNet,map_location=lambda storage, loc:storage))
else:
sys.exit("Pretrained path invalid")
@staticmethod
def classification_loss(logit, target, dataset='CelebA'):
"""Compute binary or softmax cross entropy loss."""
if dataset == 'CelebA':
return F.binary_cross_entropy_with_logits(logit, target, size_average=False) / logit.size(0)
elif dataset == 'RaFD':
return F.cross_entropy(logit, target)
def train(self,config):
train_dataset=get_loader(config.celeba_image_dir, config.attr_path,
config.selected_attrs,image_size=self.image_size,
num_workers=config.num_workers,dataset=config.dataset)
print('Start Training...')
start_time=time.time()
max_acc,epochs=0,50
for p in range(epochs):
for i,data in enumerate(train_dataset):
img, label = data
img=img.to(self.device)
label=label.to(self.device)
batch_pred = self.inc_net(img)
loss=self.classification_loss(batch_pred,label,config.dataset)
self.opt.zero_grad()
loss.backward()
self.opt.step()
if i%self.log_step==0:
et=time.time()-start_time
et = str(datetime.timedelta(seconds=et))[:-7]
acc=self.test()
print("Test Accuracy: ", acc)
if acc>max_acc:
path=os.path.join(self.save_incDir,'{}-{}-incNet.ckpt'.format(p,i))
torch.save(self.inc_net.state_dict(),path)
max_acc=acc
log = "Elapsed [{}], Epoch[{}] - Iteration [{}/{}] , loss [{}], max_acc[{}]".format(et, p,i+1,len(train_dataset),loss.item(),max_acc)
print(log)
def test(self):
acc=0
with torch.no_grad():
for _,data in enumerate(self.test_dataset):
img_test,label_test=data
label_test=label_test[:,:len(self.selected_attrs)]
img_test=img_test.to(self.device)
label_test=label_test.to(self.device)
pred=self.inc_net(img_test)
pred_label=pred>0.5
pred_label=pred_label.type(torch.FloatTensor).to(self.device)
acc+=torch.mean(torch.eq(label_test,pred_label).type(torch.FloatTensor).to(self.device)).item()
acc/=len(self.test_dataset)
return acc
@staticmethod
def flip_labels(labels,selected_attrs,dataset,hair_color_indices=None):
""" Flip trained labels randomly
Inputs:
labels: labels corresponding to image (selected_labels+n_selectedLabels)
selected_attrs: selected attributes [List]
dataset: 'CelebA' or 'RaFD'
Return:
flipped labels that the model was trained on
Shape - [batch_size,len(selected_attrs)]
"""
flipped=labels.clone()
flipped=flipped[:,:len(selected_attrs)] #discard labels that were not trained on
if dataset=='CelebA':
for i in range(len(flipped)):
if hair_color_indices is not None:
h=torch.zeros(len(hair_color_indices))
h[random.randint(0,len(hair_color_indices)-1)] =1
count=0
for j in range(len(selected_attrs)):
if hair_color_indices is not None and j in hair_color_indices:
flipped[i,j]=h[count]
count+=1
else:
flipped[i,j]=random.randint(0,1)
return flipped
def transform_op(self,image_size):
transform=[]
transform.append(T.ToPILImage())
transform.append(T.Resize(image_size))
transform.append(T.ToTensor())
transform.append(T.Normalize(mean=(0.5,0.5,0.5),std=(0.5,0.5,0.5)))
return T.Compose(transform)
def score(self,Gen):
#Inception net
self.load_pretrained()
if config.dataset=='CelebA':
hair_color_indices=[]
for i,attr_name in enumerate(self.selected_attrs):
if attr_name in ['Black_Hair', 'Blond_Hair', 'Brown_Hair', 'Gray_Hair']:
hair_color_indices.append(i)
mean_,steps=0,2
transform=self.transform_op(self.image_size)
Gen.to(device)
sigmoid=nn.Sigmoid()
data_iter=iter(test_dataset)
print("Calculating score...")
with torch.no_grad():
for i in range(steps):
try:
img, all_labels=next(data_iter)
except:
data_iter=iter(data_iter)
img,all_labels=next(data_iter) #label is a boolean labelled vector
#randomly flip
flipped_labels=self.flip_labels(all_labels,self.selected_attrs,self.dataset,hair_color_indices)
img=img.to(device)
flipped_labels=flipped_labels.to(device)
x_gen=Gen(img,flipped_labels)
x_gen=torch.stack([transform(pop.detach().cpu()) for pop in x_gen])
x_gen=x_gen.to(device)
pred_x_gen=sigmoid(self.inc_net(x_gen))
bCE=flipped_labels*torch.log(pred_x_gen)+(1-flipped_labels)*torch.log(1-pred_x_gen)
mean_+=torch.mean(torch.sum(bCE,1)).cpu().item()
print(mean_)
return mean_/steps