-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutils.py
256 lines (219 loc) · 12.1 KB
/
utils.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
#!/usr/bin/env python
import torch
from tqdm import tqdm
import os
import clip
from torch.utils.data.dataset import Dataset
from typing import Any, Callable, cast, Dict, List, Optional, Tuple
from itertools import groupby
import numpy as np
from sklearn.metrics import roc_auc_score # type: ignore
timm_models = {
'BiT_m': {'config': {'model_name': 'resnetv2_101x1_bitm', 'pretrained': True}},
'BiT_s': {'config': {'model_name': 'resnetv2_101x1_bitm',
'checkpoint_path': './model_weights/checkpoints/BiT-S-R101x1.npz'}},
'vit_base_patch16_224_21kpre': {'config': {'model_name': 'vit_base_patch16_224', 'pretrained': True}},
'vit_base_patch16_384_21kpre': {'config': {'model_name': 'vit_base_patch16_384', 'pretrained': True}},
'convnext_base_in22ft1k': {'config': {'model_name': 'convnext_base_in22ft1k', 'pretrained': True}},
'convnext_base': {'config': {'model_name': 'convnext_base', 'pretrained': True}},
'convnext_tiny-22k': {'config': {'model_name': 'convnext_tiny_384_in22ft1k', 'pretrained': True}},
'deit3_base_patch16_224': {'config': {'model_name': 'deit3_base_patch16_224', 'pretrained': True}},
'deit3_base_patch16_224_in21ft1k': {
'config': {'model_name': 'deit3_base_patch16_224_in21ft1k', 'pretrained': True}},
'tf_efficientnetv2_m': {'config': {'model_name': 'tf_efficientnetv2_m', 'pretrained': True}},
'tf_efficientnetv2_m_in21ft1k': {'config': {'model_name': 'tf_efficientnetv2_m_in21ft1k', 'pretrained': True}},
'swinv2-22k': {'config': {'model_name': 'swinv2_base_window12to16_192to256_22kft1k', 'pretrained': True}},
'swinv2-1k': {'config': {'model_name': 'swinv2_base_window16_256', 'pretrained': True}},
'deit3-384-22k': {'config': {'model_name': 'deit3_base_patch16_384_in21ft1k', 'pretrained': True}},
'deit3-384-1k': {'config': {'model_name': 'deit3_base_patch16_384', 'pretrained': True}},
'tf_efficientnet_b7_ns': {'config': {'model_name': 'tf_efficientnet_b7_ns', 'pretrained': True}},
'tf_efficientnet_b7': {'config': {'model_name': 'tf_efficientnet_b7', 'pretrained': True}},
'resnet50': {'config': {'model_name': 'resnet50', 'pretrained': True}},
'efficientnet_b0': {'config': {'model_name': 'efficientnet_b0', 'pretrained': True}},
'vit_base_patch16_384_laion2b_in12k_in1k': {
'config': {'model_name': 'vit_base_patch16_clip_384.laion2b_ft_in12k_in1k', 'pretrained': True}},
'vit_base_patch16_384_laion2b_in1k': {
'config': {'model_name': 'vit_base_patch16_clip_384.laion2b_ft_in1k', 'pretrained': True}},
'vit_base_patch16_384_openai_in12k_in1k': {
'config': {'model_name': 'vit_base_patch16_clip_384.openai_ft_in12k_in1k', 'pretrained': True}},
'vit_base_patch16_384_openai_in1k': {
'config': {'model_name': 'vit_base_patch16_clip_384.openai_ft_in1k', 'pretrained': True}},
'vit_base_patch16_384': {'config': {'model_name': 'vit_base_patch16_384.augreg_in1k', 'pretrained': True},
'batch_size': 128, 'server': 'curie'},
'xcit_medium_24_p16_224_dist': {'config': {'model_name': 'xcit_medium_24_p16_224_dist', 'pretrained': True}},
'xcit_medium_24_p16_224': {'config': {'model_name': 'xcit_medium_24_p16_224', 'pretrained': True}},
}
NINCO_class_names = ['Caracal caracal caracal', 'amphiuma_means',
'aphanizomenon_flosaquae',
'araneus_gemma',
'arctocephalus_galapagoensis',
'batrachoseps_attenuatus',
'chicken_quesadilla',
'cirsium_pitcheri', 'creme_brulee', 'ctenolepisma_longicaudata', 'cup_cakes',
'darlingtonia_californica', 'dendrolagus_lumholtzi', 'donuts', 'epithelantha_micromeris',
'erysimum_franciscanum', 'f_field_road', 'f_forest_path', 'ferocactus_pilosus', 'haemulon_sciurus',
'hippopus_hippopus', 'lasionycteris_noctivagans', 'lathyrus_odoratus', 'lepomis_auritus',
'leptoglossus_phyllopus', 'microcystis_wesenbergii', 'octopus_bimaculoides', 'octopus_rubescens',
'ozotoceros_bezoarticus', 'platycephalus_fuscus', 'polistes_dominula', 'pseudorca_crassidens',
's_sky', 'sarpa_salpa', 'sarracenia_alata', 'sepia_apama', 'sepia_officinalis',
'sepioteuthis_australis', 'spaghetti_bolognese', 'streptopus_lanceolatus', 'tapirus_bairdii',
'triturus_marmoratus', 'tursiops_aduncus', 'vaccinium_reticulatum', 'waffles',
'skipper_caterpillar']
NINCO_OOD_unit_tests_class_names = ['Gaussian', 'Rademacher', 'black', 'blobs', 'grey', 'horizontal_stripes',
'low_freq', 'low_freq_channelfullscale', 'low_freq_colorrange',
'low_freq_pixel_perm', 'monochrome', 'pixel_perm', 'tricolour', 'tricolour_primary',
'uni', 'vertical_stripes', 'white']
NINCO_popular_datasets_subsamples_class_names = [
'Places',
'iNaturalist_OOD_Plants',
'Species',
'Imagenet_O',
'OpenImage_O',
'Textures'
]
models_timm_dev = ['vit_base_patch16_384_laion2b_in12k_in1k',
'vit_base_patch16_384_laion2b_in1k',
'vit_base_patch16_384_openai_in1k',
'vit_base_patch16_384_openai_in12k_in1k',
'vit_base_patch16_384',
'xcit_medium_24_p16_224_dist',
'xcit_medium_24_p16_224']
models_timm_0_6_12 = ['BiT_m', 'BiT_s', 'vit_base_patch16_224_21kpre', 'vit_base_patch16_384_21kpre',
'convnext_base_in22ft1k', 'convnext_base', 'convnext_tiny-22k', 'deit3_base_patch16_224',
'deit3_base_patch16_224_in21ft1k', 'tf_efficientnetv2_m', 'tf_efficientnetv2_m_in21ft1k',
'swinv2-22k', 'swinv2-1k', 'deit3-384-22k', 'deit3-384-1k', 'tf_efficientnet_b7_ns',
'tf_efficientnet_b7', 'resnet50', 'efficientnet_b0', ]
models_clip = {'clip-ViT-B16': 'ViT-B/16', 'clip-ViT-L14-336': 'ViT-L/14@336px'}
def kl(p, q):
return np.sum(np.where(p != 0, p * np.log(p / q), 0))
def all_equal(iterable):
g = groupby(iterable)
return next(g, True) and not next(g, False)
def extract_features(model, dataset, savepath, wo_head=False):
torch.backends.cudnn.benchmark = True
# save
if not os.path.exists(savepath):
os.makedirs(savepath)
model.eval()
# slice dataloaders in train
slice_length = 50000
n_slices = -((-len(dataset)) // slice_length)
index_slices = {}
slice_datasets = {}
for i in range(n_slices):
# check if current iterate is already saved in directoy (for train
files = os.listdir(savepath)
complete = n_slices>1 # True for train, else False
for feat in ['logits_{}.npy'.format(i), 'features_{}.npy'.format(i), 'labels_true_{}.npy'.format(i)]:
if feat not in files:
complete = False
# if not, save create dataloader and save features
# if True: #not complete:
complete = False
if not complete:
print('Extracting features set ', i)
index_slices[i] = range(i * slice_length, min((i + 1) * slice_length, len(dataset)))
slice_datasets[i] = torch.utils.data.Subset(dataset, index_slices[i])
slice_datasets[i].__name__ = f'slice{index_slices[i].start}_to_{index_slices[i].stop}'
slice_datasets[i].classes = dataset.classes
dataloader = torch.utils.data.DataLoader(slice_datasets[i], batch_size=model.batch_size)
features = []
logits_ = []
labels_true = []
with torch.no_grad():
for (x, label) in tqdm(dataloader):
labels_true.append(label)
x = x.cuda()
feat_batch_preact = model.forward_features(x)
feat_batch = feat_batch_preact[:, 0] if wo_head else model.forward_head(feat_batch_preact,
pre_logits=True)
logits = model.forward_head(feat_batch_preact)
feat_batch = feat_batch.cpu().numpy()
logits = logits.cpu().numpy()
features.append(feat_batch)
logits_.append(logits)
# save
labels_true = torch.cat(labels_true).numpy()
features = np.concatenate(features, axis=0)
logits_ = np.concatenate(logits_, axis=0)
predictions_dict = {'logits_{}'.format(i): logits_, 'features_{}'.format(i): features,
'labels_true_{}'.format(i): labels_true}
for name, data in predictions_dict.items():
np.save(savepath + '/' + name,
data)
def extract_clip_embeddings(model, dataset, savepath, text=None, k=250):
torch.backends.cudnn.benchmark = True
# save
if not os.path.exists(savepath):
os.makedirs(savepath)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=model.batch_size)
features = []
labels_true = []
model.eval()
with torch.no_grad():
for i, (x, label) in tqdm(enumerate(dataloader)):
labels_true.append(label)
x = x.cuda()
feat_batch = model.encode_image(x)
feat_batch = feat_batch.cpu().numpy()
features.append(feat_batch)
if (i + 1) % k == 0: # save every k batches
labels_true = torch.cat(labels_true).numpy()
features = np.concatenate(features, axis=0)
predictions_dict = {'features_{}'.format(i): features,
'labels_true_{}'.format(i): labels_true}
for name, data in predictions_dict.items():
np.save(savepath + '/' + name,
data) # with open(savepath + '/' + name, 'wb') as f: # pickle.dump(features, f)
features = []
labels_true = []
# save remaining
labels_true = torch.cat(labels_true).numpy()
features = np.concatenate(features, axis=0)
predictions_dict = {'features_{}'.format(0): features, 'labels_true_{}'.format(0): labels_true}
# text embeddings of labels
if text is not None:
text_inputs = torch.cat([clip.tokenize(f"a photo of a {c}") for c in text]).cuda()
with torch.no_grad():
text_encoded = model.encode_text(text_inputs)
predictions_dict['text_encoded_0'] = text_encoded.cpu().numpy()
#
for name, data in predictions_dict.items():
np.save(savepath + '/' + name,
data)
def auroc_ood(values_in: np.ndarray, values_out: np.ndarray) -> float:
"""
Implementation of Area-under-Curve metric for out-of-distribution detection.
The higher the value the better.
Args:
values_in: Maximal confidences (i.e. maximum probability per each sample)
for in-domain data.
values_out: Maximal confidences (i.e. maximum probability per each sample)
for out-of-domain data.
Returns:
Area-under-curve score.
"""
if len(values_in) * len(values_out) == 0:
return np.NAN
y_true = len(values_in) * [1] + len(values_out) * [0]
y_score = np.nan_to_num(np.concatenate([values_in, values_out]).flatten())
return roc_auc_score(y_true, y_score)
def fpr_at_tpr(values_in: np.ndarray, values_out: np.ndarray, tpr: float) -> float:
"""
Calculates the FPR at a particular TRP for out-of-distribution detection.
The lower the value the better.
Args:
values_in: Maximal confidences (i.e. maximum probability per each sample)
for in-domain data.
values_out: Maximal confidences (i.e. maximum probability per each sample)
for out-of-domain data.
tpr: (1 - true positive rate), for which probability threshold is calculated for
in-domain data.
Returns:
False positive rate on out-of-domain data at (1 - tpr) threshold.
"""
if len(values_in) * len(values_out) == 0:
return np.NAN
t = np.quantile(values_in, (1 - tpr))
fpr = (values_out >= t).mean()
return fpr