File size: 11,749 Bytes
cf167af |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 |
#!/usr/bin/env python
import argparse
import json
from os.path import basename, splitext
import os
import mmengine
import numpy as np
import pandas as pd
import torch
from numpy.linalg import norm, pinv
from scipy.special import logsumexp, softmax
from sklearn import metrics
from sklearn.covariance import EmpiricalCovariance
from sklearn.metrics import pairwise_distances_argmin_min
from tqdm import tqdm
import pickle
from os.path import dirname
import torchvision as tv
from PIL import Image
from mmpretrain.apis import init_model
def parse_args():
parser = argparse.ArgumentParser(description='Detect an image')
parser.add_argument(
'--cfg', help='Path to config',
default='/dataset/jingyaoli/AD/MOOD_/MOODv2/configs/beit-base-p16_224px.py')
parser.add_argument('--ood_feature',
default=None, help='Path to ood feature file')
parser.add_argument(
'--checkpoint', help='Path to checkpoint',
default='/dataset/jingyaoli/AD/MOODv2/pretrain/beit-base_3rdparty_in1k_20221114-c0a4df23.pth',)
parser.add_argument('--img_path', help='Path to image',
default='/dataset/jingyaoli/AD/MOOD_/MOODv2/imgs/DTD_cracked_0004.jpg')
parser.add_argument('--fc',
default='/dataset/jingyaoli/AD/MOODv2/outputs/beit-224px/fc.pkl', help='Path to fc path')
parser.add_argument('--id_data', default='imagenet', help='id data name')
parser.add_argument('--id_train_feature',
default='/dataset/jingyaoli/AD/MOODv2/outputs/beit-224px/imagenet_train.pkl', help='Path to data')
parser.add_argument('--id_val_feature',
default='/dataset/jingyaoli/AD/MOODv2/outputs/beit-224px/imagenet_test.pkl', help='Path to output file')
parser.add_argument('--ood_features',
default=None, nargs='+', help='Path to ood features')
parser.add_argument(
'--methods', nargs='+',
default=['MSP', 'MaxLogit', 'Energy', 'Energy+React', 'ViM', 'Residual', 'GradNorm', 'Mahalanobis', ], # 'KL-Matching'
help='methods')
parser.add_argument(
'--train_label',
default='datalists/imagenet2012_train_random_200k.txt',
help='Path to train labels')
parser.add_argument(
'--clip_quantile', default=0.99, help='Clip quantile to react')
parser.add_argument(
'--fpr', default=95, help='False Positive Rate')
return parser.parse_args()
def evaluate(method, score_id, score_ood, target_fpr):
threhold = np.percentile(score_id, 100 - target_fpr)
if score_ood >= threhold:
print('\033[94m', method, '\033[0m', 'evaluation:', '\033[92m', 'in-distribution', '\033[0m')
else:
print('\033[94m', method, '\033[0m', 'evaluation:', '\033[91m', 'out-of-distribution', '\033[0m')
def kl(p, q):
return np.sum(np.where(p != 0, p * np.log(p / q), 0))
def gradnorm(x, w, b, num_cls):
fc = torch.nn.Linear(*w.shape[::-1])
fc.weight.data[...] = torch.from_numpy(w)
fc.bias.data[...] = torch.from_numpy(b)
fc.cuda()
x = torch.from_numpy(x).float().cuda()
logsoftmax = torch.nn.LogSoftmax(dim=-1).cuda()
confs = []
for i in tqdm(x, desc='Computing Gradnorm ID/OOD score'):
targets = torch.ones((1, num_cls)).cuda()
fc.zero_grad()
loss = torch.mean(
torch.sum(-targets * logsoftmax(fc(i[None])), dim=-1))
loss.backward()
layer_grad_norm = torch.sum(torch.abs(
fc.weight.grad.data)).cpu().numpy()
confs.append(layer_grad_norm)
return np.array(confs)
def extract_image_feature(args):
torch.backends.cudnn.benchmark = True
print('=> Loading model')
cfg = mmengine.Config.fromfile(args.cfg)
model = init_model(cfg, args.checkpoint, 0).cuda().eval()
print('=> Loading image')
if hasattr(cfg.model.backbone, 'img_size'):
img_size = cfg.model.backbone.img_size
else:
img_size = 224
transform = tv.transforms.Compose([
tv.transforms.Resize((img_size, img_size)),
tv.transforms.ToTensor(),
tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
x = transform(Image.open(args.img_path).convert('RGB')).unsqueeze(0)
print('=> Extracting feature')
with torch.no_grad():
x = x.cuda()
if cfg.model.backbone.type == 'BEiTPretrainViT':
# (B, L, C) -> (B, C)
feat_batch = model.backbone(
x, mask=None)[0].mean(1)
elif cfg.model.backbone.type == 'SwinTransformer':
# (B, C, H, W) -> (B, C)
feat_batch = model.backbone(x)[0]
B, C, H, W = feat_batch.shape
feat_batch = feat_batch.reshape(B, C, -1).mean(-1)
else:
# (B, C)
feat_batch = model.backbone(x)[0]
assert len(feat_batch.shape) == 2
feature = feat_batch.cpu().numpy()
print(f'Extracted Feature: {feature.shape}')
return feature
def main():
args = parse_args()
if args.ood_feature and os.path.exists(args.ood_feature):
feature_ood = mmengine.load(args.ood_feature)
else:
feature_ood = extract_image_feature(args)
if os.path.exists(args.fc):
w, b = mmengine.load(args.fc)
print(f'{w.shape=}, {b.shape=}')
num_cls = len(b)
train_labels = np.array([
int(line.rsplit(' ', 1)[-1])
for line in mmengine.list_from_file(args.train_label)
], dtype=int)
print(f'image path: {args.img_path}')
print('=> Loading features')
feature_id_train = mmengine.load(args.id_train_feature).squeeze()
feature_id_val = mmengine.load(args.id_val_feature).squeeze()
print(f'{feature_id_train.shape=}, {feature_id_val.shape=}')
if os.path.exists(args.fc):
print('=> Computing logits...')
logit_id_train = feature_id_train @ w.T + b
logit_id_val = feature_id_val @ w.T + b
logit_ood = feature_ood @ w.T + b
print('=> Computing softmax...')
softmax_id_train = softmax(logit_id_train, axis=-1)
softmax_id_val = softmax(logit_id_val, axis=-1)
softmax_ood = softmax(logit_ood, axis=-1)
u = -np.matmul(pinv(w), b)
# ---------------------------------------
method = 'MSP'
if method in args.methods:
score_id = softmax_id_val.max(axis=-1)
score_ood = softmax_ood.max(axis=-1)
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'MaxLogit'
if method in args.methods:
score_id = logit_id_val.max(axis=-1)
score_ood = logit_ood.max(axis=-1)
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'Energy'
if method in args.methods:
score_id = logsumexp(logit_id_val, axis=-1)
score_ood = logsumexp(logit_ood, axis=-1)
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'Energy+React'
if method in args.methods:
clip = np.quantile(feature_id_train, args.clip_quantile)
logit_id_val_clip = np.clip(
feature_id_val, a_min=None, a_max=clip) @ w.T + b
score_id = logsumexp(logit_id_val_clip, axis=-1)
logit_ood_clip = np.clip(feature_ood, a_min=None, a_max=clip) @ w.T + b
score_ood = logsumexp(logit_ood_clip, axis=-1)
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'ViM'
if method in args.methods:
if feature_id_val.shape[-1] >= 2048:
DIM = num_cls
elif feature_id_val.shape[-1] >= 768:
DIM = 512
else:
DIM = feature_id_val.shape[-1] // 2
ec = EmpiricalCovariance(assume_centered=True)
ec.fit(feature_id_train - u)
eig_vals, eigen_vectors = np.linalg.eig(ec.covariance_)
NS = np.ascontiguousarray(
(eigen_vectors.T[np.argsort(eig_vals * -1)[DIM:]]).T)
vlogit_id_train = norm(np.matmul(feature_id_train - u, NS), axis=-1)
alpha = logit_id_train.max(axis=-1).mean() / vlogit_id_train.mean()
vlogit_id_val = norm(np.matmul(feature_id_val - u, NS), axis=-1) * alpha
energy_id_val = logsumexp(logit_id_val, axis=-1)
score_id = -vlogit_id_val + energy_id_val
energy_ood = logsumexp(logit_ood, axis=-1)
vlogit_ood = norm(np.matmul(feature_ood - u, NS), axis=-1) * alpha
score_ood = -vlogit_ood + energy_ood
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'Residual'
if method in args.methods:
if feature_id_val.shape[-1] >= 2048:
DIM = 1000
elif feature_id_val.shape[-1] >= 768:
DIM = 512
else:
DIM = feature_id_val.shape[-1] // 2
ec = EmpiricalCovariance(assume_centered=True)
ec.fit(feature_id_train - u)
eig_vals, eigen_vectors = np.linalg.eig(ec.covariance_)
NS = np.ascontiguousarray(
(eigen_vectors.T[np.argsort(eig_vals * -1)[DIM:]]).T)
score_id = -norm(np.matmul(feature_id_val - u, NS), axis=-1)
score_ood = -norm(np.matmul(feature_ood - u, NS), axis=-1)
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'GradNorm'
if method in args.methods:
score_ood = gradnorm(feature_ood, w, b, num_cls)
score_id = gradnorm(feature_id_val, w, b, num_cls)
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'Mahalanobis'
if method in args.methods:
train_means = []
train_feat_centered = []
for i in tqdm(range(train_labels.max() + 1), desc='Computing classwise mean feature'):
fs = feature_id_train[train_labels == i]
_m = fs.mean(axis=0)
train_means.append(_m)
train_feat_centered.extend(fs - _m)
ec = EmpiricalCovariance(assume_centered=True)
ec.fit(np.array(train_feat_centered).astype(np.float64))
mean = torch.from_numpy(np.array(train_means)).cuda().float()
prec = torch.from_numpy(ec.precision_).cuda().float()
score_id = -np.array(
[(((f - mean) @ prec) * (f - mean)).sum(axis=-1).min().cpu().item()
for f in tqdm(torch.from_numpy(feature_id_val).cuda().float(), desc='Computing Mahalanobis ID score')])
score_ood = -np.array([
(((f - mean) @ prec) * (f - mean)).sum(axis=-1).min().cpu().item()
for f in tqdm(torch.from_numpy(feature_ood).cuda().float(), desc='Computing Mahalanobis OOD score')
])
result = evaluate(method, score_id, score_ood, args.fpr)
# ---------------------------------------
method = 'KL-Matching'
if method in args.methods:
pred_labels_train = np.argmax(softmax_id_train, axis=-1)
mean_softmax_train = []
for i in tqdm(range(num_cls), desc='Computing classwise mean softmax'):
mean_softmax = softmax_id_train[pred_labels_train == i]
if mean_softmax.shape[0] == 0:
mean_softmax_train.append(np.zeros((num_cls)))
else:
mean_softmax_train.append(np.mean(mean_softmax, axis=0))
score_id = -pairwise_distances_argmin_min(
softmax_id_val, np.array(mean_softmax_train), metric=kl)[1]
score_ood = -pairwise_distances_argmin_min(
softmax_ood, np.array(mean_softmax_train), metric=kl)[1]
result = evaluate(method, score_id, score_ood, args.fpr)
if __name__ == '__main__':
main()
|