-
Notifications
You must be signed in to change notification settings - Fork 3
/
postprocessing.py
317 lines (250 loc) · 9.58 KB
/
postprocessing.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
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
307
308
309
310
311
312
313
314
315
316
317
import os
import yaml
import copy
import math
import random
import argparse
import itertools
import numpy as np
import os.path as osp
import matplotlib.pyplot as plt
from PIL import Image
from time import time
from os import path as osp
from multiprocessing import Pool
import torch
from torch.utils import data
from torch.nn import functional as F
from torchvision import transforms
from torchvision import datasets
from torchvision import models
# Diffusion Model imports
from guided_diffusion import dist_util
from guided_diffusion.script_util import (
model_and_diffusion_defaults,
diffusion_defaults,
create_model_and_diffusion,
create_gaussian_diffusion,
create_classifier,
args_to_dict,
add_dict_to_argparser,
)
from guided_diffusion.sample_utils import (
get_DiME_iterative_sampling,
clean_class_cond_fn,
dist_cond_fn,
ImageSaver,
SlowSingleLabel,
load_from_DDP_model,
PerceptualLoss,
ChunkedDataset,
)
from guided_diffusion.gaussian_diffusion import _extract_into_tensor
from guided_diffusion.image_datasets import BINARYDATASET, MULTICLASSDATASETS
# core imports
from core.utils import print_dict, merge_all_chunks
from core.metrics import accuracy, loic_flip, get_prediction
from core.attacks_and_models import Normalizer, joint_classifier_ddpm, get_attack
# model imports
from models import get_classifier
import matplotlib
matplotlib.use('Agg') # to disable display
# =======================================================
# =======================================================
# Functions
# =======================================================
# =======================================================
def create_args():
defaults = dict(
clip_denoised=True,
batch_size=16,
gpu='0',
image_size=256,
# path args
model_path='pretrained_models/256x256_diffusion_uncond.pt',
classifier_path='',
output_path='',
exp_name='',
label_query=1,
label_target=-1,
seed=4,
dataset='',
# filtering args
sampling_time_fraction=0.1, # sampling fraction
sampling_stochastic=True,
sampling_inpaint=0.0,
sampling_dilation=15,
)
defaults.update(model_and_diffusion_defaults())
parser = argparse.ArgumentParser()
add_dict_to_argparser(parser, defaults)
return parser.parse_args()
# =======================================================
# =======================================================
# Custom functions
# =======================================================
# =======================================================
@torch.inference_mode()
def filter_fn(
diffusion,
model,
steps,
x,
orig,
mask,
device,
stochastic,
inpaint, # set a thresh t, 0 < t < 1 as the inpaint. those values in m_i < t will be used as repaint
):
indices = list(range(steps))[::-1]
x = (x - 0.5) / 0.5
orig = (orig - 0.5) / 0.5
noise_fn = torch.randn_like if stochastic else torch.zeros_like
boolmask = (mask < inpaint).float()
for idx, t in enumerate(indices):
# filter the with the diffusion model
t = torch.tensor([t] * x.size(0), device=x.device)
if idx == 0:
x = diffusion.q_sample(x, t, noise=noise_fn(x))
if inpaint != 0:
x = (x * (1 - boolmask) +
boolmask * diffusion.q_sample(orig, t, noise=noise_fn(x)))
out = diffusion.p_mean_variance(
model, x, t,
clip_denoised=True
)
x = out['mean']
if stochastic and (idx != (steps - 1)):
x += torch.exp(0.5 * out["log_variance"]) * torch.randn_like(x)
if inpaint != 0:
x = x * (1 - boolmask) + boolmask * orig
x = (x * 0.5) + 0.5
x = x.clamp(0, 1)
return x
class CFDataset():
def __init__(self, output_path, exp_name, folder_name='all_filtered', dilation=15):
self.path = output_path
self.exp_name = exp_name
self.images = []
self.dilation = dilation
self.output_path = osp.join(self.path, 'Results', self.exp_name, folder_name)
for cc, ccf in itertools.product(['CC', 'IC'], ['CCF', 'ICF']):
os.makedirs(osp.join(self.output_path, cc, ccf, 'CF'), exist_ok=True)
os.makedirs(osp.join(self.output_path, cc, ccf, 'SM'), exist_ok=True)
self.images += [(cc, ccf, f) for f in os.listdir(osp.join(output_path, 'Results', exp_name, 'attack', cc, ccf, 'CF'))]
def __len__(self):
return len(self.images)
def __getitem__(self, idx):
cc, ccf, f = self.images[idx]
image = self._load_images(osp.join(self.path, 'Original', 'Correct' if cc == 'CC' else 'Incorrect', f))
image = self.transform(image)
cf = self._load_images(osp.join(self.path, 'Results', self.exp_name, 'attack', cc, ccf, 'CF', f))
cf = self.transform(cf)
# generates mask
delta = (cf - image).abs().sum(dim=0, keepdim=True)
delta /= delta.max()
delta = F.max_pool2d(delta, self.dilation, stride=1, padding=(self.dilation - 1) // 2)
return image, cf, delta, cc, ccf, f
@staticmethod
def _load_images(path):
with open(path, "rb") as f:
img = Image.open(f)
return img.convert("RGB")
@staticmethod
def transform(img):
img = np.array(img)
img = img.astype(np.float32) / 255
img = img.transpose(2, 0, 1) # C x H x W
img = torch.from_numpy(img).float()
return img
# =======================================================
# =======================================================
# Main
# =======================================================
# =======================================================
def main():
args = create_args()
normal_steps = int(args.sampling_time_fraction * int(args.timestep_respacing if args.timestep_respacing != '' else args.diffusion_steps))
print(args)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
os.makedirs(osp.join(args.output_path, 'Results'),
exist_ok=True)
# ========================================
# Set seeds
# ========================================
torch.manual_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
# ========================================
# Load Dataset
# ========================================
dataset = CFDataset(args.output_path, args.exp_name,
folder_name='all_filtered_r-{}_i-{}_d-{}_f-{}'.format(args.timestep_respacing, args.sampling_inpaint, args.sampling_dilation, args.sampling_time_fraction))
print('Images on the dataset:', len(dataset))
loader = data.DataLoader(dataset, batch_size=args.batch_size,
shuffle=False,
num_workers=4, pin_memory=True)
# ========================================
# load models
# ========================================
print('Loading Model and diffusion model')
# respaced diffusion has the respaced strategy
model, diffusion = create_model_and_diffusion(
**args_to_dict(args, model_and_diffusion_defaults().keys())
)
model.load_state_dict(
dist_util.load_state_dict(args.model_path, map_location="cpu")
)
model.to(dist_util.dev())
if args.use_fp16:
model.convert_to_fp16()
model.eval()
print('Loading Classifier')
classifier = get_classifier(args)
classifier.to(dist_util.dev()).eval()
# ========================================
# get custom function for the forward phase
# and other variables of interest
# ========================================
start_time = time()
print('Starting Image Generation')
for idx, (img, cf, masks, CC, _, files) in enumerate(loader):
print(f'Iter {idx} / {len(loader)} | Time: {int(time() - start_time)}s', end='\r')
cf = cf.to(dist_util.dev())
img = img.to(dist_util.dev())
masks = masks.to(dist_util.dev())
# sample image from the noisy_img
filt = filter_fn(
diffusion=diffusion,
model=model,
steps=normal_steps,
x=cf,
orig=img,
mask=masks,
device=dist_util.dev(),
stochastic=args.sampling_stochastic,
inpaint=args.sampling_inpaint,
)
# Initial Classification, no noise included
_, c_pred = get_prediction(classifier, img, args.dataset in BINARYDATASET)
_, f_pred = get_prediction(classifier, filt, args.dataset in BINARYDATASET)
CCF = []
if args.dataset in BINARYDATASET:
looper = (c_pred != f_pred)
else:
correct = (c_pred == args.label_query).float()
looper = (args.label_target == f_pred) * correct + (args.label_query == f_pred) * (1 - correct)
for c in looper:
CCF.append('CCF' if c.item() == 1 else 'ICF')
filt = (filt * 255).permute((0, 2, 3, 1)).cpu().numpy().astype('uint8')
masks = ((masks < args.sampling_inpaint) * 255).permute((0, 2, 3, 1)).squeeze(3).cpu().numpy().astype('uint8')
# save images
for fi, ma, cc, ccf, f in zip(filt, masks, CC, CCF, files):
Image.fromarray(fi).save(osp.join(dataset.output_path, cc, ccf, 'CF', f))
Image.fromarray(ma).save(osp.join(dataset.output_path, cc, ccf, 'SM', f))
if (idx + 1) == len(loader):
print(f'Iter {idx + 1} / {len(loader)} | Time: {int(time() - start_time)}s')
print('\nDone')
break
if __name__ == '__main__':
main()