-
Notifications
You must be signed in to change notification settings - Fork 1
/
main.py
226 lines (198 loc) · 12.1 KB
/
main.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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
import sys
import copy
import os
import argparse
import numpy as np
import torch
import torchvision
from torch.utils.data import TensorDataset, DataLoader, TensorDataset
from torchsummary import summary
from torch.utils.data import DataLoader, TensorDataset
import matplotlib.pyplot as plt
############################
from TDRA import datasets, models, utils, trainers
from setting import args_parser
if __name__ == '__main__':
args = args_parser() ## Reading the input arguments (see setting.py)
if torch.cuda.is_available():
args.device = "cuda"
args.gpu = 0
torch.cuda.set_device(args.gpu)
# ### Optional: For making your results reproducible
# torch.manual_seed(args.rand_seed)
# torch.cuda.manual_seed(args.rand_seed)
# torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = False
# np.random.seed(args.rand_seed)
## Fetch the datasets
if args.dataset == "mnist":
(train_images, train_labels), (valid_images, valid_labels), (test_images, test_labels) = datasets.get_dataset(args)
train_dataset = ((train_images, train_labels), (valid_images, valid_labels))
if args.dataset == "f_mnist":
(train_images, train_labels), (valid_images, valid_labels), (test_images, test_labels) = datasets.get_dataset(args)
train_dataset = ((train_images, train_labels), (valid_images, valid_labels))
if args.dataset == "cifar10":
(train_images, train_labels), (valid_images, valid_labels), (test_images, test_labels) = datasets.get_dataset(args)
train_dataset = ((train_images, train_labels), (valid_images, valid_labels))
if args.dataset == "cifar100":
(train_images, train_labels), (valid_images, valid_labels), (test_images, test_labels) = datasets.get_dataset(args)
train_dataset = ((train_images, train_labels), (valid_images, valid_labels))
if args.dataset == "tinyimgnet":
(train_images, train_labels), (valid_images, valid_labels), (test_images, test_labels) = datasets.get_dataset(args)
train_dataset = ((train_images, train_labels), (valid_images, valid_labels))
elif args.dataset == "celeba":
(train_images, train_labels), (valid_images, valid_labels), (test_images, test_labels) = datasets.get_dataset(args)
train_labels, valid_labels, test_labels = datasets.prepare_labels(args, train_labels, test_labels, valid_labels)
train_dataset = ((train_images, train_labels), (valid_images, valid_labels))
pos_weights = torch.Tensor((len(train_labels) - train_labels.sum(0))/train_labels.sum(0))
# pass
print("\n*** Dataset's Info")
print("Training")
utils.print_dataset_info(args, (train_images, train_labels))
# if args.dataset == "celeba" or args.dataset == "mnist":
print("Validation")
utils.print_dataset_info(args, (valid_images, valid_labels))
print("Testing")
utils.print_dataset_info(args, (test_images, test_labels))
## For logging
exp_name = args.dataset+"_"+args.output+"/"+\
str(len(args.attributes))+"_"+str(int(args.beta_1*100))+"_"+str(int(args.beta_2*100))+\
"_"+str(args.image_size)+"_"+str(int(args.alpha_ssim*100))+"_"+str(int(args.alpha_mse*100))+\
"_"+str(args.with_aug)+"_"+str(args.model_width)+"_"+str(args.rand_seed)
## Models
if args.dataset == "mnist":
model_F = models.WideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.MNIST_MEAN,
std = datasets.MNIST_STD,
num_input_channels=args.n_channels)
model_G = models.DecWideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.MNIST_MEAN,
std = datasets.MNIST_STD,
num_input_channels=args.n_channels,
input_type = args.output)
if args.dataset == "f_mnist":
model_F = models.WideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.F_MNIST_MEAN,
std = datasets.F_MNIST_STD,
num_input_channels=args.n_channels)
model_G = models.DecWideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.F_MNIST_MEAN,
std = datasets.F_MNIST_STD,
num_input_channels=args.n_channels,
input_type = args.output)
elif args.dataset == "cifar10":
model_F = models.WideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.CIFAR10_MEAN,
std = datasets.CIFAR10_STD,
num_input_channels=args.n_channels)
model_G = models.DecWideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.CIFAR10_MEAN,
std = datasets.CIFAR10_STD,
num_input_channels=args.n_channels,
input_type = args.output)
elif args.dataset == "cifar100":
model_F = models.WideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.CIFAR100_MEAN,
std = datasets.CIFAR100_STD,
num_input_channels=args.n_channels)
model_G = models.DecWideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.CIFAR100_MEAN,
std = datasets.CIFAR100_STD,
num_input_channels=args.n_channels,
input_type = args.output)
elif args.dataset == "tinyimgnet":
model_F = models.WideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.TINYIMGNET_MEAN,
std = datasets.TINYIMGNET_STD,
num_input_channels=args.n_channels)
model_G = models.DecWideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.TINYIMGNET_MEAN,
std = datasets.TINYIMGNET_STD,
num_input_channels=args.n_channels,
input_type = args.output)
elif args.dataset == "celeba":
model_F = models.WideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.CELEBA_MEAN,
std = datasets.CELEBA_STD,
num_input_channels=args.n_channels)
model_G = models.DecWideResNet(num_classes = len(args.attributes),
width = args.model_width,
mean = datasets.CELEBA_MEAN,
std = datasets.CELEBA_STD,
num_input_channels=args.n_channels,
input_type = args.output)
model_F.to(args.device)
model_G.to(args.device)
summary(model_F, input_size=(args.n_channels, args.image_size, args.image_size), device=args.device)
summary(model_G, input_size=(len(args.attributes),), device=args.device)
save_dir = args.root_dir+"/results/"+exp_name+"/"
if not os.path.exists(save_dir):
os.makedirs(save_dir)
if args.dataset == "celeba":
model_F, model_G = trainers.train_enc_dec_cat(args, model_F, model_G, train_dataset, pos_weights, save_path=save_dir)
else:
model_F, model_G = trainers.train_enc_dec_cat(args, model_F, model_G, train_dataset, pos_weights=None, save_path=save_dir)
model_F.load_state_dict(torch.load(save_dir+"best_model_F.pt", map_location=torch.device(args.device)))
model_G.load_state_dict(torch.load(save_dir+"best_model_G.pt", map_location=torch.device(args.device)))
test_data_loader = DataLoader(TensorDataset(torch.Tensor(test_images), torch.Tensor(test_labels).long()),
batch_size=args.n_batch, shuffle=False, drop_last=False)
mse_err, ssim_err, att_accs = utils.evaluate_acc_enc_dec_cat(args, model_F, model_G, test_data_loader)
print("Test MSE: {:.4f}".format(mse_err))
d_range = test_images.max()-test_images.min()
PSNR = 10.0 * np.log10(d_range**2/mse_err)
print("Test PSNR: {:.2f}".format(PSNR))
print("Test SSIM: {:.4f}".format(1-ssim_err))
print("Test Accs:", att_accs*100)
print("Mean Test Accs:", np.mean(att_accs)*100)
def compute_meanD_and_covM(train_images, valid_images):
data = np.append(train_images, valid_images, axis=0)
meanD = data.mean(0).copy()
data = data.reshape(data.shape[0], np.prod(data.shape[1:]))
print(data.shape)
covM = np.cov(data, rowvar=False)
return meanD, covM
meanD, covM = compute_meanD_and_covM(train_images, valid_images)
meanD = torch.Tensor(meanD).to(args.device)
covM = torch.Tensor(covM).to(args.device)
def compute_mah_dist(x, y, inv_covM):
diff = x - y
# torch.sqrt(torch.dot(diff[0], torch.matmul(inv_covM, diff[0]))))
return torch.sqrt(torch.einsum('nj,jk,nk->n', diff, inv_covM, diff))
def compute_rec_risk(args, model_F, model_G, dataloader, meanD, covM):
model_F.eval()
model_G.eval()
n_samples = 0
rec_risk = 0.
inv_covM = torch.linalg.pinv(covM)
for batch_id, (images, labels) in enumerate(dataloader):
with torch.no_grad():
images, labels = images.to(args.device), labels.to(args.device)
n_samples += len(labels)
logits = model_F(images)
rec_images = model_G(logits)
org_x = torch.reshape(images, (images.shape[0],np.prod(images.shape[1:])))
rec_x = torch.reshape(rec_images, (rec_images.shape[0],np.prod(rec_images.shape[1:])))
mean_x = torch.reshape(meanD, (np.prod(meanD.shape),))
# utils.imshow(torchvision.utils.make_grid(images[:10], nrow=10))
# utils.imshow(torchvision.utils.make_grid(rec_images[:10], nrow=10))
# utils.imshow(torchvision.utils.make_grid(meanD, nrow=10))
md_xr = compute_mah_dist(org_x, rec_x, inv_covM)
md_xm = compute_mah_dist(org_x, mean_x, inv_covM)
rec_risk += torch.div(md_xm, md_xr)
rec_risk = torch.sum(rec_risk)/n_samples
return rec_risk
test_data_loader = DataLoader(TensorDataset(torch.Tensor(test_images), torch.Tensor(test_labels).long()),
batch_size=args.n_batch, shuffle=False, drop_last=True)
rec_risk = compute_rec_risk(args, model_F, model_G, test_data_loader, meanD, covM)
print("# Epsilon: {:.4f}".format(rec_risk.item()))