-
Notifications
You must be signed in to change notification settings - Fork 0
/
generation_v3.py
408 lines (311 loc) · 14.6 KB
/
generation_v3.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
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
# Generation V1: con regularización y null bg
import os
import cv2
import sys
import time
import scipy
import torch
import argparse
import numpy as np
import torch.optim
import shutil
from formal_utils import *
from skimage.transform import resize
from PIL import ImageFilter, Image
import matplotlib.pyplot as plt
from skimage.transform import resize
from torchvision import models
from skimage.util import random_noise
use_cuda = torch.cuda.is_available()
# Fixing for deterministic results
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def numpy_to_torch(img, requires_grad=True):
if len(img.shape) < 3:
output = np.float32([img])
else:
output = np.float32(np.transpose(img.copy(), (2, 0, 1)))
output_t = torch.from_numpy(output)
if use_cuda:
output_t = output_t.to('cuda') # cuda()
output_t.unsqueeze_(0)
output_t.requires_grad = requires_grad
return output_t
def numpy_to_torch2(img):
if len(img.shape) < 3:
output = np.float32([img])
else:
output = np.transpose(img, (2, 0, 1))
output = torch.from_numpy(output)
output.unsqueeze_(0)
return output
if __name__ == '__main__':
# img_path = 'perro_gato.jpg'
img_path = 'dog.jpg'
# img_path = 'example.JPEG'
# img_path = 'example_2.JPEG'
# img_path = 'goldfish.jpg'
save_path = './output/'
# gt_category = 207 # Golden retriever
# gt_category = 281 # tabby cat
gt_category = 258 # "Samoyed, Samoyede"
# gt_category = 282 # tigger cat
# gt_category = 565 # freight car
#gt_category = 1 # goldfish
try:
shutil.rmtree(save_path)
except OSError as e:
print("Error: %s : %s" % (save_path, e.strerror))
# PyTorch random seed
torch.manual_seed(0)
learning_rate = 0.1 # 0.1 (preservation sparser) 0.3 (preservation dense)
max_iterations = 101
l1_coeff = 1e-6 #2*0.5*1e-7 # 1e-4 (preservation)
size = 224
tv_beta = 3
tv_coeff = 1e-2
factorTV = 0 # 1(dense) o 0.5 (sparser/sharp) #0.5 (preservation)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# model = models.vgg16(pretrained=True)
# model = models.resnet50(pretrained=True)
model = models.googlenet(pretrained=True)
model.to(device)
# evaluar el modelo para que sea deterministico
model.eval()
list_of_layers = ['conv1',
'conv2',
'conv3',
'inception3a',
'inception3b',
'inception4a',
'inception4b',
'inception4c',
'inception4d',
'inception4e',
'inception5a',
'inception5b',
'fc'
]
label_map = load_imagenet_label_map()
# model = torch.nn.DataParallel(model).to('cuda')
# model = model.to('cuda')
activation_orig = {}
gradients_orig = {}
# no se necesitan gradientes para los parametros
# for param in model.parameters():
# param.requires_grad = False
def get_activation_orig(name):
def hook(model, input, output):
activation_orig[name] = output.clone()
return hook
def get_gradients_orig(name):
def hook(model, grad_input, grad_output):
gradients_orig[name] = grad_output[0].cpu().detach().numpy()
return hook
for name, layer in model.named_children():
if name in list_of_layers:
F_hook = layer.register_forward_hook(get_activation_orig(name))
B_hook = layer.register_backward_hook(get_gradients_orig(name))
init_time = time.time()
# Leer la imágen del archivo
# original_img = cv2.imread(img_path, 1)
# img = np.float32(original_img) / 255
original_img_pil = Image.open(img_path).convert('RGB')
# voy a agregar un poco de ruido a la imagen
#noise_img = Image.effect_noise((500, 375), 25)
#noise_img_3 = Image.merge('RGB', (noise_img, noise_img, noise_img))
#img_sum = Image.blend(original_img_pil, noise_img_3, 0.5)
#original_np = np.array(img_sum)
#plt.imshow(original_np)
#plt.show()
# normalización de acuerdo al promedio y desviación std de Imagenet
transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
# se normaliza la imágen y se agrega una dimensión [1,3,244,244]
img_normal = transform(original_img_pil).unsqueeze(0) # Tensor (1, 3, 224, 224)
img_normal.requires_grad = False
img_normal = img_normal.to(device)
cat_orig = label_map[gt_category]
print('explicacion para: ', cat_orig)
# Path to the output folder
save_path = os.path.join(save_path, 'MP', 'imagenet')
mkdir_p(os.path.join(save_path))
# Compute original output
# org_softmax = torch.nn.Softmax(dim=1)(model(preprocess_image(img, size)))
org_softmax = torch.nn.Softmax(dim=1)(model(img_normal)) # tensor(1,1000)
prob_orig = org_softmax.data[0, gt_category].cpu().detach().numpy()
o_img_path = os.path.join(save_path, 'real_{}_{:.3f}_image.jpg'
.format(cat_orig.split(',')[0].split(' ')[0].split('-')[0], prob_orig))
# visualización de tensor normalizado a array y desrnomalizado
img_transform_T = np.moveaxis(img_normal[0, :].cpu().detach().numpy().transpose(), 0, 1) # array (224,224,3)
img_unormalize = np.uint8(255 * unnormalize(img_transform_T)) # array (224,224,3)
Image.fromarray(img_unormalize).save(o_img_path, 'JPEG')
print('probabilidad original para ', cat_orig, '=', prob_orig)
F_hook.remove()
B_hook.remove()
del model
# CALCULO ITERATIVO DE LA MASCARA
model = models.googlenet(pretrained=True)
model.to(device)
model.eval()
gradients = {}
def get_activation_mask(name):
def hook(model, input, output):
act_mask = output
# print(act_mask.shape). #debug
# print(activation_orig[name].shape) #debug
limite_sup = (act_mask <= torch.fmax(torch.tensor(0), activation_orig[name]))
limite_inf = (act_mask >= torch.fmin(torch.tensor(0), activation_orig[name]))
oper = limite_sup * limite_inf
# print('oper shape=',oper.shape). #debug
act_mask.requires_grad_(True)
act_mask.retain_grad()
h = act_mask.register_hook(lambda grad: grad * oper)
# x.register_hook(update_gradients(2))
# activation[name]=act_mask
# h.remove()
return hook
def get_act_mask_gradients(name):
def hook(model, grad_input, grad_output):
gradients[name] = grad_output[0]
# print('backward')
# return (new_grad,)
return hook
for name, layer in model.named_children():
if name in list_of_layers:
layer.register_forward_hook(get_activation_mask(name))
layer.register_backward_hook(get_act_mask_gradients(name))
for param in model.parameters():
param.requires_grad = True
img = img_normal # tensor (1, 3, 224, 224)
np.random.seed(seed=0)
# mask = np.random.uniform(0, 0.01, size=(224, 224)) # array (224, 224) generation
# mask = np.random.rand(224, 224)
mask = np.random.uniform(0.99, 1, size=(224, 224)) # array (224, 224) preservation
mask = numpy_to_torch(mask) # tensor (1, 1, 224, 224)
#null_img = torch.zeros(1, 3, size, size).to(device) # tensor (1, 3, 224, 224)
# imagen nulla difuminada
orig_img_blur = original_img_pil.filter(ImageFilter.GaussianBlur(10))
null_img_blur = transform(orig_img_blur).unsqueeze(0)
null_img_blur.requires_grad = False
null_img = null_img_blur.to(device)
# Definición del tipo de optimizador
optimizer = torch.optim.Adam([mask], lr=learning_rate)
# optimizer = torch.optim.SGD([mask], lr=learning_rate, momentum = 0.9)
# momentum = 0.9
# optimizer = torch.optim.SGD([mask],
# lr=learning_rate,
# momentum=momentum,
# dampening=momentum)
for i in range(max_iterations):
# upsampled_mask = upsample(mask)
# The single channel mask is used with an RGB image,
# so the mask is duplicated to have 3 channel,
upsampled_mask = mask.expand(1, 3, mask.size(2), mask.size(3)) # tensor (1, 3, 224, 224)
# upsampled_mask = mask
perturbated_input = img.mul(upsampled_mask) + null_img.mul(1 - upsampled_mask)
optimizer.zero_grad()
outputs = torch.nn.Softmax(dim=1)(model(perturbated_input)) # tensor (1, 1000)
similarity = -(org_softmax.data[0, gt_category] * torch.log(outputs[0, gt_category])) # tensor
# + tv_coeff * tv_norm(mask, tv_beta)
# loss = l1_coeff * torch.sum(torch.abs(mask)) + similarity + factorTV * tv_coeff * tv_norm(mask,
# tv_beta) # tensor
#loss = l1_coeff * torch.sum(torch.abs(1 - mask)) + outputs[0, gt_category]
# loss = l1_coeff * torch.sum(torch.abs(mask)) + similarity
loss = l1_coeff * torch.sum(torch.abs(mask)) - torch.log(outputs[0, gt_category])
loss.backward()
# mask_grads=np.squeeze(mask.grad.data.cpu().numpy())
# print('max mask(grad)=', mask_grads.max())
# print('min mask(grad)=', mask_grads.min())
# torch.nn.utils.clip_grad_norm_(mask, 1, norm_type=float('inf'))
# mask.grad.data = torch.nn.functional.normalize(mask.grad.data, p=2, dim=(2, 3))
# torch.nn.utils.clip_grad_norm_(mask, 1)
# mask_grads = np.squeeze(mask.grad.data.cpu().numpy())
# print('max mask(grad) after clip=', mask_grads.max())
# print('min mask(grad) after clip=', mask_grads.min())
optimizer.step()
mask.data.clamp_(0, 1) # mask tensor (1, 1, 224, 224)
# debug visualización de la mascara
# mask_np = np.squeeze(mask.cpu().detach().numpy()) # array fp32 (224, 224)
# plt.imshow(1 - mask_np) # 1-mask para deletion
# plt.title("mask value")
# plt.show()
# debug visualización del gradiente de la mask
# maskgrads_np = np.squeeze(mask.grad.data.cpu().numpy())
# plt.imshow(maskgrads_np)
# plt.title("mask grad")
# plt.show()
# control
# upsampled_mask_control = mask.expand(1, 3, mask.size(2), mask.size(3)) # tensor (1, 3, 224, 224)
# up_mask_np = upsampled_mask_control.cpu().detach().numpy()
# Create save_path for storing intermediate steps
path = os.path.join(save_path, 'intermediate_steps')
mkdir_p(path)
# DEBUG
# mask_np = np.squeeze(mask.cpu().detach().numpy()) # array fp32 (224, 224)
# print('max mask=', mask_np.max())
# print('min mask=', mask_np.min())
# mask_grads=np.squeeze(mask.grad.data.cpu().numpy())
# print('max mask(grad)=', mask_grads.max())
# print('min mask(grad)=', mask_grads.min())
# torch.nn.utils.clip_grad_norm_(mask, 1)
# mask_grads = np.squeeze(mask.grad.data.cpu().numpy())
# print('max mask(grad) after clip=', mask_grads.max())
# print('min mask(grad) after clip=', mask_grads.min())
# plt.imshow(mask_np)
# plt.show()
# DEBUG
if (i % 20) == 0:
# Save intermediate steps
amax, aind = outputs.max(dim=1)
gt_val = outputs.data[:, gt_category]
img_pert_np = perturbated_input[0, :].cpu().detach().numpy() # array (3, 224, 224)
img_pert_np_T = img_pert_np.transpose() # array (224, 224, 3)
img_pert_np_T2 = np.moveaxis(img_pert_np_T, 0, 1) # array (224, 224, 3) se intercambian cols 0 y 1
img_pert_unnorma = np.uint8(255 * unnormalize(img_pert_np_T2)) # array enteros (224, 224, 3)
path_intermediate = os.path.abspath(os.path.join(path, 'intermediate_{:05d}_{}_{:.3f}_{}_{:.3f}.jpg'
.format(i,
label_map[aind.item()].split(',')[0].split(' ')[
0].split('-')[0],
amax.item(),
label_map[gt_category].split(',')[0].split(' ')[
0].split('-')[0],
gt_val.item())))
Image.fromarray(img_pert_unnorma).save(path_intermediate, 'JPEG')
# np.save(os.path.abspath(os.path.join(save_path, "mask_MP.npy")),
# 1 - mask.cpu().detach().numpy()[0, 0, :])
# up_mask_np = upsampled_mask.cpu().detach().numpy()
# plt.imshow(up_mask_np[0, 0, :])
# plt.show()
print('prediccion:', outputs[0, gt_category].cpu().detach().numpy())
mask_np = np.squeeze(mask.cpu().detach().numpy()) # array fp32 (224, 224)
# mask_np_T = np.moveaxis(mask_np.transpose(), 0, 1)
print('max mask=', mask_np.max())
print('min mask=', mask_np.min())
plt.imshow(mask_np) # 1-mask para deletion
plt.show()
print('Time taken: {:.3f}'.format(time.time() - init_time))
original_img_pil = Image.open(img_path).convert('RGB')
img_normal = transform(original_img_pil).unsqueeze(0) # Tensor (1, 3, 224, 224)
mask_tensor = numpy_to_torch2(1-mask_np) # tensor (1, 1, 224, 224)
mask_expanded = mask_tensor.expand(1, 3, mask.size(2), mask.size(3)) # tensor (1, 3, 224, 224)
null_img = torch.zeros(1, 3, size, size)
img_masked = img_normal.mul(mask_expanded)
# transforma de (PIL o tensor) de (1,3,224,224) a np; desnormaliza y grafica
img_normal_np = img_masked.numpy()
img_transform_T = np.moveaxis(img_normal_np[0, :].transpose(), 0, 1)
img_unormalize = np.uint8(255 * unnormalize(img_transform_T))
plt.imshow(img_unormalize)
plt.show()
# img_normal2 = transform(Image.fromarray(img_pert_unnorma)).unsqueeze(0) # array -> PIL y retorna Tensor (1, 3, 224, 224)
# img_normal2_np = img_normal2.numpy()
# plt.imshow(img_pert_unnorma)
# plt.show()
org_softmax = torch.nn.Softmax(dim=1)(model(img_masked.to(device)))
prob_orig = org_softmax.data[0, gt_category].cpu().detach().numpy()
print('probabilidad de la mascara complemento=', prob_orig)