-
Notifications
You must be signed in to change notification settings - Fork 0
/
somethingIdontKnowWhyItWorks.py
430 lines (341 loc) · 13.7 KB
/
somethingIdontKnowWhyItWorks.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
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import matplotlib.pyplot as plt
import numpy as np
from tqdm import tqdm
from torchvision import datasets, transforms
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import multiprocessing
import numpy as np
import os
import sklearn.utils.class_weight
# for reproducibility
torch.manual_seed(1)
VEC_PATH = './embeddings/'
IMG_PATH = None
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if device == 'cuda':
torch.backends.cudnn.benchmark = True
NUM_PROCESS = 8
l = 10
m = 5
r = 512
def unpacking_apply_along_axis(all_args):
(func1d, axis, arr, args, kwargs) = all_args
"""
Like numpy.apply_along_axis(), but with arguments in a tuple
instead.
This function is useful with multiprocessing.Pool().map(): (1)
map() only handles functions that take a single argument, and (2)
this function can generally be imported from a module, as required
by map().
"""
return np.apply_along_axis(func1d, axis, arr, *args, **kwargs)
def parallel_apply_along_axis(func1d, axis, arr, *args, **kwargs):
effective_axis = 1 if axis == 0 else axis
if effective_axis != axis:
arr = arr.swapaxes(axis, effective_axis)
# Chunks for the mapping (only a few chunks):
chunks = [(func1d, effective_axis, sub_arr, args, kwargs)
for sub_arr in np.array_split(arr, NUM_PROCESS)]
# print(chunks)
pool = multiprocessing.Pool(processes=NUM_PROCESS)
individual_results = pool.map(unpacking_apply_along_axis, chunks)
# Freeing the workers:
pool.close()
pool.join()
return np.concatenate(individual_results)
def parallel_matrix_operation(func, arr):
# chunks = [(func, sub_arr) for sub_arr in np.array_split(arr, NUM_PROCESS)]
chunks = np.array_split(arr, NUM_PROCESS)
pool = multiprocessing.Pool(processes=NUM_PROCESS)
individual_results = pool.map(func, chunks)
pool.close()
pool.join()
return np.concatenate(individual_results)
def check_accuracy(loader, model):
num_correct = 0
num_samples = 0
model.eval()
with torch.no_grad():
for x, y in loader:
x = x.to(device=device)
y = y.to(device=device)
scores = model(x)
_, predictions = scores.max(1)
num_correct += (predictions == y).sum()
num_samples += predictions.size(0)
print(f'Got {num_correct} / {num_samples} with accuracy {float(num_correct) / float(num_samples) * 100:.2f}')
def float_to_binary(x, m, n):
x_abs = np.abs(x)
x_scaled = round(x_abs * 2 ** n)
res = '{:0{}b}'.format(x_scaled, m + n)
if x >= 0:
res = '0' + res
else:
res = '1' + res
return res
# binary to float
def binary_to_float(bstr, m, n):
sign = bstr[0]
bs = bstr[1:]
res = int(bs, 2) / 2 ** n
if int(sign) == 1:
res = -1 * res
return res
def string_to_int(a):
bit_str = "".join(x for x in a)
return np.array(list(bit_str)).astype(int)
def join_string(a, num_bit=l, num_feat=r):
res = np.empty(num_feat, dtype="S10")
# res = []
for i in range(num_feat):
# res.append("".join(str(x) for x in a[i*l:(i+1)*l]))
res[i] = "".join(str(x) for x in a[i * l:(i + 1) * l])
return res
def float_bin(x):
return float_to_binary(x, m, l - m - 1)
def bin_float(x):
return binary_to_float(x, m, l - m - 1)
def BitRand(sample_feature_arr, eps=10.0, l=10, m=5):
r = sample_feature_arr.shape[1]
float_to_binary_vec = np.vectorize(float_bin)
binary_to_float_vec = np.vectorize(bin_float)
feat_tmp = parallel_matrix_operation(float_to_binary_vec, sample_feature_arr)
feat = parallel_apply_along_axis(string_to_int, axis=1, arr=feat_tmp)
rl = r * l
sum_ = 0
for k in range(l):
sum_ += np.exp(2 * eps * k / l)
alpha = np.sqrt((eps + rl) / (2 * r * sum_))
index_matrix = np.array(range(l))
index_matrix = np.tile(index_matrix, (sample_feature_arr.shape[0], r))
p = 1 / (1 + alpha * np.exp(index_matrix * eps / l))
p_temp = np.random.rand(p.shape[0], p.shape[1])
perturb = (p_temp > p).astype(int)
perturb_feat = (perturb + feat) % 2
perturb_feat = parallel_apply_along_axis(join_string, axis=1, arr=perturb_feat)
# print(perturb_feat)
return torch.tensor(parallel_matrix_operation(binary_to_float_vec, perturb_feat), dtype=torch.float)
class AMIADatasetCelebA(Dataset):
def __init__(self, target, transform, dataroot, train=True, imgroot=None, multiplier=100):
self.target = target
self.target_multiplier = multiplier
self.transform = transform
if train:
self.valid_data = np.arange(162770, 182637)
self.length = len(target) * multiplier + len(self.valid_data)
else:
self.train_data = np.arange(62770)
mask = np.ones(62770, dtype=bool)
mask[target] = False
self.train_data = self.train_data[mask, ...]
self.length = len(self.train_data) + len(target) * multiplier
self.dataroot = dataroot
self.imgroot = imgroot
self.data_name = sorted(os.listdir(dataroot))
self.train = train
def __len__(self):
return self.length
def __getitem__(self, idx):
if self.train == False:
if idx / self.target_multiplier < len(self.target):
filename = self.data_name[self.target[int(idx / self.target_multiplier)]]
# img_loc = os.path.join(self.dataroot, self.data_name[self.target[idx]])
class_id = torch.tensor(int(idx / self.target_multiplier))
else:
idx -= len(self.target) * self.target_multiplier
filename = self.data_name[self.train_data[idx]]
# img_loc = os.path.join(self.dataroot, self.data_name[self.valid_data[idx]])
class_id = torch.tensor(len(self.target))
else:
if idx / self.target_multiplier < len(self.target):
filename = self.data_name[self.target[int(idx / self.target_multiplier)]]
# img_loc = os.path.join(self.dataroot, self.data_name[self.target[idx]])
class_id = torch.tensor(int(idx / self.target_multiplier))
else:
idx -= len(self.target) * self.target_multiplier
filename = self.data_name[self.valid_data[idx]]
# img_loc = os.path.join(self.dataroot, self.data_name[self.valid_data[idx]])
class_id = torch.tensor(len(self.target))
if self.imgroot:
img = Image.open(self.imgroot + filename)
img = self.transform(img)
else:
img = torch.tensor([])
# img_tensor = img2vec.get_vec(img, tensor=True)
# img_tensor = torch.squeeze(img_tensor)
img_tensor = torch.load(self.dataroot + filename)
# img_tensor = img_tensor + s1.astype(np.float32)
return img_tensor, class_id, img
class Classifier(nn.Module):
def __init__(self, n_inputs, n_outputs):
super(Classifier, self).__init__()
self.fc1 = nn.Linear(n_inputs, 1000)
self.fc2 = nn.Linear(1000, n_outputs)
def forward(self, x):
x = torch.flatten(x, 1)
fc1 = self.fc1(x)
fc1 = F.relu(fc1)
fc2 = self.fc2(fc1)
out = torch.sigmoid(fc2)
probs = F.softmax(out, dim=1)
return out, probs, fc2, fc1
def tpr_tnr(prediction, truth):
""" Returns the confusion matrix for the values in the `prediction` and `truth`
tensors, i.e. the amount of positions where the values of `prediction`
and `truth` are
- 1 and 1 (True Positive)
- 1 and 0 (False Positive)
- 0 and 0 (True Negative)
- 0 and 1 (False Negative)
"""
confusion_vector = prediction / truth
# Element-wise division of the 2 tensors returns a new tensor which holds a
# unique value for each case:
# 1 where prediction and truth are 1 (True Negative)
# inf where prediction is 1 and truth is 0 (False Negative)
# nan where prediction and truth are 0 (True Positive)
# 0 where prediction is 0 and truth is 1 (False Positive)
true_negatives = torch.sum(confusion_vector == 1).item()
false_negatives = torch.sum(confusion_vector == float('inf')).item()
true_positives = torch.sum(torch.isnan(confusion_vector)).item()
false_positives = torch.sum(confusion_vector == 0).item()
# print(true_negatives, false_negatives, true_positives, false_positives)
return true_positives / (true_positives + false_negatives), true_negatives / (true_negatives + false_positives), (
true_positives + true_negatives) / (true_negatives + false_negatives + true_positives + false_positives)
num_target = 1
target = [0]
transform = transforms.Compose([
transforms.Resize(64),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
train_loader = torch.utils.data.DataLoader(
AMIADatasetCelebA(target, transform, VEC_PATH, True, imgroot=IMG_PATH, multiplier=4000), shuffle=False,
num_workers=0, batch_size=200000)
test_loader = torch.utils.data.DataLoader(
AMIADatasetCelebA(target, transform, VEC_PATH, False, imgroot=None, multiplier=5000), shuffle=False, num_workers=0,
batch_size=200000)
eps = 3.0
if num_target == 1:
SAVE_NAME = f'CELEBA_embed_BitRand_single_{target[0]}_{eps}.pth'
else:
SAVE_NAME = f'CELEBA_embed_BitRand_multiple_{num_target}_{eps}.pth'
np.random.seed(1)
x_train, y_train, imgs_train = next(iter(train_loader))
x_train[1:] = BitRand(x_train[1:], eps)
print("Start generate positive")
temp = x_train.numpy()
positive = []
negative = []
for i in tqdm(range(temp.shape[0])):
if i < 4000:
indx_pos = list(range(4000))
indx_pos.remove(i)
idx_pos = np.random.choice(indx_pos, size=1, replace=False)[0]
positive.append(temp[idx_pos, :].tolist())
indx_neg = list(range(4000,temp.shape[0]))
idx_neg = np.random.choice(indx_neg, size=1, replace=False)[0]
negative.append(temp[idx_neg, :].tolist())
else:
indx_pos = list(range(4000, temp.shape[0]))
indx_pos.remove(i)
idx_pos = np.random.choice(indx_pos, size=1, replace=False)[0]
positive.append(temp[idx_pos, :].tolist())
indx_neg = list(range(4000))
idx_neg = np.random.choice(indx_neg, size=1, replace=False)[0]
negative.append(temp[idx_neg, :].tolist())
anchor = torch.from_numpy(temp)
positive = torch.Tensor(positive)
negative = torch.Tensor(negative)
print(anchor.size(), positive.size(), negative.size())
x_test, y_test, _ = next(iter(test_loader))
x_test[1:] = BitRand(x_test[1:], eps)
anchor = anchor.to(device)
positive = positive.to(device)
negative = negative.to(device)
y_train = y_train.to(device)
x_test = x_test.to(device)
y_test = y_test.to(device)
y_test_threat = torch.cat((y_test[:1], y_test[-50:]))
x_test_threat = torch.cat((x_test[:1], x_test[-50:]))
weight = sklearn.utils.class_weight.compute_class_weight('balanced', classes=np.arange(num_target + 1), y=y_test.cpu().detach().numpy())
model = Classifier(x_train.shape[1], num_target + 1)
model = model.to(device)
if device == 'cuda':
model = torch.nn.DataParallel(model)
custom_weight = np.array([1600.0, 200.0])
criterion = nn.CrossEntropyLoss(weight=torch.tensor(custom_weight, dtype=torch.float).to(device))
triplet_loss = nn.TripletMarginLoss(margin=1.0, p=2)
min_loss = 100000000000
max_correct = 0
max_tpr = 0.0
max_tnr = 0.0
epoch = 0
lr = 1e-6
optimizer = optim.Adam(model.parameters(), lr=lr)
# optimizer = optim.Adam(model.parameters(), lr=lr)
from tqdm import tqdm
for i in range(4000):
num_correct = 0
num_samples = 0
loss_value = 0
epoch += 1
# for imgs, labels in iter(train_loader):
model.train()
out_anchor, probs_anchor, fc2_anchor, fc1_anchor = model(anchor)
_, _, _, fc1_positive = model(positive)
_, _, _, fc1_negative = model(negative)
loss = criterion(out_anchor, y_train) + triplet_loss(fc1_anchor, fc1_positive, fc1_negative)
loss_value += loss
predictions = fc2_anchor[:, 0] < 0
tpr_train, tnr_train, _ = tpr_tnr(predictions, y_train)
loss.backward()
optimizer.step() # make the updates for each parameter
optimizer.zero_grad() # a clean up step for PyTorch
# Test acc
out, probs, fc2, _ = model(x_test_threat)
predictions = fc2[:, 0] < 0
tpr, tnr, _ = tpr_tnr(predictions, y_test_threat)
acc = (tpr + tnr) / 2
if (tpr + tnr) / 2 > max_tpr:
state = {
'net': model.state_dict(),
'test': (tpr, tnr),
'train': (tpr_train, tnr_train),
'acc': acc,
'lr': lr,
'epoch': epoch
}
max_tpr = (tpr + tnr) / 2
torch.save(state, SAVE_NAME)
if i % 1000 == 0:
# print(f'Loss: {loss_value.item()} | Acc: {num_correct}/{num_samples} | Epoch: {i}')
print(
f'Loss: {loss_value.item()} | Train_TPR = {tpr_train}, Train_TNR = {tnr_train:.5f} | TPR = {tpr}, TNR = {tnr}, ACC = {acc} | Epoch: {epoch}')
print(torch.load(SAVE_NAME)['train'])
print(torch.load(SAVE_NAME)['test'])
print(torch.load(SAVE_NAME)['acc'])
print(torch.load(SAVE_NAME)['epoch'])
D = 20
tpr = 0
for i in range(5000):
x_test_threat = torch.cat((x_test[i:i + 1], x_test[np.random.randint(5000, 64769, D - 1)]))
out, probs, fc2, _ = model(x_test_threat)
if torch.sum(fc2[:, 0] > 0) > 0:
tpr += 1
tpr /= 5000
print(f'tpr = {tpr}')
tnr = 0
for i in range(5000):
x_test_threat = x_test[np.random.randint(5000, 64769, D)]
out, probs, fc2, _ = model(x_test_threat)
if torch.sum(fc2[:, 0] > 0) == 0:
tnr += 1
tnr /= 5000
print(f'tnr = {tnr}')
print(f'adv = {tpr / 2 + tnr / 2}')