-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtest_osr_ood_TTA.py
182 lines (145 loc) · 8.02 KB
/
test_osr_ood_TTA.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
import os
import torch
import torchvision.transforms
import data.dataset_osr_test
import model.get_model
import utils.test_option
import numpy as np
import torch.nn as nn
from copy import deepcopy
from tqdm import tqdm
from torch.optim.swa_utils import AveragedModel
from torch.utils.data import DataLoader
from torch.autograd import Variable
IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
def get_target(dataset_root, num_classes=1000):
cls_num = np.zeros(num_classes, dtype=int)
class_folders = [folder for folder in os.listdir(dataset_root) if os.path.isdir(os.path.join(dataset_root, folder))]
for class_index, folder in enumerate(class_folders):
class_path = os.path.join(dataset_root, folder)
num_images = len([img for img in os.listdir(class_path) if os.path.isfile(os.path.join(class_path, img))])
cls_num[class_index] = num_images
target = cls_num / np.sum(cls_num)
return target
@torch.no_grad()
def test_predict_GradNorm_RP(model, test_loader, targets, num_classes=1000):
"""
Get class predictions and Grad Norm Score for all instances in loader
"""
model.eval()
id_preds = [] # Store class preds
gradnorm_preds = [] # Stores OSR preds
save_labels = []
image_names = []
targets = torch.tensor(targets).cuda()
targets = targets.unsqueeze(0)
feat_model = deepcopy(model)
feat_model.module.head = nn.Sequential()
# First extract all features
for b,(images, labels, _, filenames) in enumerate(tqdm(test_loader)):
inputs = Variable(images.cuda(), requires_grad=True)
# Get logits
features = feat_model(inputs)
outputs = model.module.head.forward(features)
U = torch.norm(features, p=1, dim=1)
out_softmax = torch.nn.functional.softmax(outputs, dim=1)
V = torch.norm((targets - out_softmax), p=1, dim=1)
S = U * V / num_classes
# id_preds.extend(out_softmax.argmax(dim=-1).detach().cpu().numpy())
id_preds.extend(out_softmax.detach().cpu().numpy())
gradnorm_preds.extend(S.detach().cpu().numpy())
save_labels.extend(labels.detach().cpu().numpy())
image_names.extend(filenames)
id_preds = np.array(id_preds)
gradnorm_preds = np.array(gradnorm_preds)
save_labels = np.array(save_labels)
return id_preds, gradnorm_preds, save_labels, image_names
def get_tencrop_dataload(imagenet_1k_root, imagenet_21k_root, batch_size, input_size, idx, resize_ratio=0.875, data_json_path="./splits/imagenet_ssb_splits.json"):
transform = torchvision.transforms.Compose([
torchvision.transforms.Resize(int(input_size/resize_ratio)),
torchvision.transforms.TenCrop(input_size),
torchvision.transforms.Lambda(lambda crops: crops[idx]),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=torch.tensor(IMAGENET_DEFAULT_MEAN), std=torch.tensor(IMAGENET_DEFAULT_STD))])
datasets = data.dataset_osr_test.get_imagenet_osr_test_datasets(test_transform=transform,
imagenet_1k_root=imagenet_1k_root,
imagenet_21k_root=imagenet_21k_root,
data_path=data_json_path)
dataloaders = {}
for k, v, in datasets.items():
if v is not None:
dataloaders[k] = DataLoader(v, batch_size, pin_memory=True, drop_last=False, sampler=None, num_workers=4)
return dataloaders['test_known'], dataloaders['test_unknown']
def get_fivecrop_dataload(imagenet_1k_root, imagenet_21k_root, batch_size, input_size, idx, resize_ratio=0.875, data_json_path="./splits/imagenet_ssb_splits.json"):
transform = torchvision.transforms.Compose([
torchvision.transforms.Resize(int(input_size/resize_ratio)),
torchvision.transforms.FiveCrop(input_size),
torchvision.transforms.Lambda(lambda crops: crops[idx]),
# torchvision.transforms.ColorJitter(brightness=0.5, hue=0.3),
# torchvision.transforms.RandomHorizontalFlip(1.0),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=torch.tensor(IMAGENET_DEFAULT_MEAN), std=torch.tensor(IMAGENET_DEFAULT_STD))])
datasets = data.dataset_osr_test.get_imagenet_osr_test_datasets(test_transform=transform,
imagenet_1k_root=imagenet_1k_root,
imagenet_21k_root=imagenet_21k_root,
data_path=data_json_path)
dataloaders = {}
for k, v, in datasets.items():
if v is not None:
dataloaders[k] = DataLoader(v, batch_size, pin_memory=True, drop_last=False, sampler=None, num_workers=4)
return dataloaders['test_known'], dataloaders['test_unknown']
if __name__ == "__main__":
args = utils.test_option.get_args_parser()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
torch.manual_seed(100)
iter_num_dict = {
"10c" : 10,
"5c" : 5,
}
mode = args.mode
preds_mode = "GradNorm_RP"
iter_num = iter_num_dict[mode]
save_dir =args.save_dir
save_dir = os.path.join(save_dir, mode)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
model_path = args.save_pth
net = model.get_model.get_model(1000, args)
if args.optim_name == 'fmfp' or args.optim_name == 'swa':
net = AveragedModel(net)
net.load_state_dict(torch.load(model_path), strict=True)
net = net.cuda()
net.module.norm = nn.Identity()
imagenet_1k_root = r"/datassd/Inet1K/"
imagenet_1k_train = os.path.join(imagenet_1k_root, "train")
imagenet_21k_root = r"/data/ImageNet-21K/"
data_json_path = r"./splits/imagenet_ssb_splits.json"
batch_size = args.batch_size
input_size = args.input_size
crop_resize_ratio = args.crop_ratio
print(f"crop_resize_ratio: {crop_resize_ratio}")
for idx in range(iter_num):
if mode == "5c":
dataloader_id, dataloader_ood = get_fivecrop_dataload(imagenet_1k_root, imagenet_21k_root, batch_size, input_size, idx, resize_ratio=crop_resize_ratio, data_json_path=data_json_path)
elif mode == "10c":
dataloader_id, dataloader_ood = get_tencrop_dataload(imagenet_1k_root, imagenet_21k_root, batch_size, input_size, idx, resize_ratio=crop_resize_ratio, data_json_path=data_json_path)
print(len(dataloader_id.dataset))
print(len(dataloader_ood.dataset))
save_sub_dir = os.path.join(save_dir, str(idx))
if not os.path.exists(save_sub_dir):
os.makedirs(save_sub_dir)
target = get_target(imagenet_1k_train)
id_preds, osr_preds_id_samples, id_labels, id_image_names = test_predict_GradNorm_RP(net, dataloader_id, target)
_, osr_preds_osr_samples, _, ood_image_names = test_predict_GradNorm_RP(net, dataloader_ood, target)
np.save(os.path.join(save_sub_dir, "id_preds_softmax.npy"), id_preds)
np.save(os.path.join(save_sub_dir, "id_preds_labels.npy"), id_labels)
np.save(os.path.join(save_sub_dir, "id_preds_score.npy"), osr_preds_id_samples)
np.save(os.path.join(save_sub_dir, "ood_preds_score.npy"), osr_preds_osr_samples)
id_image_names_w = [name + "\n" for name in id_image_names]
with open(os.path.join(save_sub_dir, "id_image_name.txt"), 'w') as fd:
fd.writelines(id_image_names_w)
ood_image_names_w = [name + "\n" for name in ood_image_names]
with open(os.path.join(save_sub_dir, "ood_image_name.txt"), 'w') as fd:
fd.writelines(ood_image_names_w)
print(f"save predict {idx} Successful!")