-
Notifications
You must be signed in to change notification settings - Fork 2
/
eval_model.py
256 lines (197 loc) · 8.84 KB
/
eval_model.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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
import argparse
import os, sys
import numpy as np
import pandas as pd
import pickle as pk
import torch
import torch.nn as nn
import torchvision.datasets as datasets
import torchvision.transforms as transforms
import torchvision.models as models
from torch.autograd import grad
parser = argparse.ArgumentParser('Gathers statistics of an ImageNet model and writes these stats into a DataFrame.')
parser.add_argument('data-dir', type=str,
metavar='DIR',
help='Directory where ImageNet data is saved')
parser.add_argument('--model', type=str, default='resnet152',
choices=['resnet152'], help='Model')
parser.add_argument('--batch-size', type=int, default=100,metavar='N',
help='number of images to evaluate at a time')
parser.add_argument('--dataset',type=str, choices=['coco','imagenet'])
parser.add_argument('--attack-path', type=str, default=None)
def main():
args = parser.parse_args()
if args.attack_path is None:
args.save_path = os.path.join('./logs/', args.dataset, args.model, 'eval.pkl')
else:
d, f = os.path.split(args.attack_path)
name = os.path.splitext(f)[0]
args.save_path = os.path.join(d, name+'-eval.pkl')
pth = os.path.split(args.save_path)[0]
os.makedirs(pth, exist_ok=True)
print('Arguments:')
for p in vars(args).items():
print(' ',p[0]+': ',p[1])
print('\n')
has_cuda = torch.cuda.is_available()
# Data loading code
if args.dataset=='imagenet':
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if args.attack_path is None:
traindir = os.path.join(args.data_dir, 'train')
valdir = os.path.join(args.data_dir, 'val')
loader = torch.utils.data.DataLoader(
datasets.ImageFolder(valdir, transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size, shuffle=False,
num_workers=4, pin_memory=True)
else:
with open(args.attack_path,'rb') as fo:
d, f = os.path.split(args.attack_path)
name, ext = os.path.splitext(f)
if ext=='.pkl':
dct = pk.load(fo)
elif ext=='.npz':
dct=np.load(fo)
x = dct['perturbed']
y = dct['labels']
y = torch.from_numpy(y)
x = torch.from_numpy(x)
test = torch.utils.data.TensorDataset(x,y)
loader = torch.utils.data.DataLoader(test,
batch_size=args.batch_size,
num_workers=4,
shuffle=False,
pin_memory=has_cuda)
Nsamples = len(loader.dataset)
Nclasses = 1000
m = getattr(models, args.model)(pretrained=True).cuda()
elif args.dataset == 'coco':
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
Nclasses = 80
Nsamples = 5000
traindir = os.path.join(args.data_dir, 'images/train/train2014')
valdir = os.path.join(args.data_dir, 'images/val/val2014')
annotation_file = os.path.join(args.data_dir, 'annotations/instances_val2014.json')
dset = datasets.CocoDetection(valdir, annotation_file, transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
]))
sub_idx = np.random.choice(np.arange(len(dset)), size=Nsamples)
subset = torch.utils.data.Subset(dset, sub_idx)
loader = torch.utils.data.DataLoader(subset,
batch_size=args.batch_size, shuffle=False,
num_workers=1, pin_memory=True)
if args.attack_path is not None:
raise NotImplementedError
m = getattr(models, args.model)(pretrained=True).cuda()
m.eval()
for p in m.parameters():
p.requires_grad_(False)
if torch.cuda.device_count()>1:
m = nn.DataParallel(m)
class GradNorms(nn.Module):
def __init__(self, norm, summary=lambda x: x, create_graph=False, retain_graph=False):
super().__init__()
self.norm = norm
self.create_graph=create_graph
self.summary = summary
self.retain_graph = retain_graph
def forward(self, l, x):
sh = x.shape
bsz = sh[0]
if not x.requires_grad:
x.requires_grad_()
dx, = grad(l, x, create_graph=self.create_graph, retain_graph=self.retain_graph)
dx = dx.view(bsz, -1)
if self.norm in [2,'2']:
n = dx.norm(p=2,dim=-1)
elif self.norm in [1,'1']:
n = dx.norm(p=1,dim=-1)
elif self.norm in ['inf',inf]:
n = dx.abs().max(dim=-1)[0]
else:
raise ValueError('%s is not an available norm'%self.norm)
return self.summary(n)
GradFunc1 = GradNorms(2,retain_graph=False)
criterion = nn.CrossEntropyLoss(reduction='none').cuda()
Loss = torch.zeros(Nsamples).cuda()
Top1 = torch.zeros(Nsamples,dtype=torch.uint8).cuda()
Rank = torch.zeros(Nsamples,dtype=torch.int64).cuda()
Top5 = torch.zeros(Nsamples,dtype=torch.uint8).cuda()
ModelSqGradx = torch.zeros(Nsamples).cuda()
NegLogPmax = torch.zeros(Nsamples).cuda()
NegLogP5 = torch.zeros(Nsamples).cuda()
Entropy = torch.zeros(Nsamples).cuda()
sys.stdout.write('\nRunning through dataloader:\n')
Nc = 1000 if args.model=='resnet152' else Nclasses
Jx = torch.arange(Nc).cuda().view(1,-1)
Jx = Jx.expand(args.batch_size, Nc)
for i, (x,y) in enumerate(loader):
sys.stdout.write(' Completed [%6.2f%%]\r'%(100*i*args.batch_size/Nsamples))
sys.stdout.flush()
if args.dataset != 'coco':
x, y = x.cuda(), y.cuda()
else:
x = x.cuda()
try:
y = y[0]['category_id']
y = y.cuda()
except (KeyError, IndexError): # hack for now
y = torch.zeros(args.batch_size, dtype=torch.int64).cuda()
x.requires_grad_(True)
yhat = m(x)
p = yhat.softmax(dim=-1)
e = (-p*p.log()).sum(dim=-1)
vs, js = yhat.sort(dim=-1,descending=True)
b = js==y.view(-1,1)
rank = Jx[b]
pmax = p.max(dim=-1)[0]
log = pmax.log()
p5 = p.topk(5,dim=-1)[0]
sump5 = p.sum(dim=-1)
pnorm = p.norm(dim=-1)
loss = criterion(yhat, y)
dpn = GradFunc1(pnorm.sum(),x)
t5 = p.topk(5,dim=-1)[0]
t1 = t5[:,0]
top1 = torch.argmax(yhat,dim=-1)==y
s = yhat.sort(dim=-1, descending=True)[1]
top5 = (s[:,0:5]==y.view(args.batch_size,1)).sum(dim=-1)
ix = torch.arange(i*args.batch_size, (i+1)*args.batch_size,device=x.device)
Loss[ix] = loss.detach()
Rank[ix]= rank.detach()
Top1[ix] = top1.detach()
Top5[ix] = top5.detach().type(torch.uint8)
ModelSqGradx[ix] = dpn.detach()
NegLogPmax[ix] = -log.detach()
NegLogP5[ix]= -sump5.log()
Entropy[ix] = e.detach()
sys.stdout.write(' Completed [%6.2f%%]\r'%(100.))
df = pd.DataFrame({'loss':Loss.cpu().numpy(),
'top1':np.array(Top1.cpu().numpy(),dtype=np.bool),
'top5':np.array(Top5.cpu().numpy(), dtype=np.bool),
'gradx_modelsq_2norm': ModelSqGradx.cpu().numpy(),
'neg_log_pmax': NegLogPmax.cpu().numpy(),
'neg_log_p5': NegLogP5.cpu().numpy(),
'model_entropy': Entropy.cpu().numpy(),
'rank': Rank.cpu().numpy()})
ix1 = np.array(df['top1'], dtype=bool)
ix5 = np.array(df['top5'], dtype=bool)
ix15 = np.logical_or(ix5,ix1)
ixw = np.logical_not(np.logical_or(ix1, ix5))
df['type'] = pd.DataFrame(ix1.astype(np.int8) + ix5.astype(np.int8))
d = {0:'mis-classified',1:'top5',2:'top1'}
df['type'] = df['type'].map(d)
df['type'] = df['type'].astype('category')
df.to_pickle(args.save_path)
if __name__=='__main__':
main()