-
Notifications
You must be signed in to change notification settings - Fork 0
/
mnist_mlp_urn.py
409 lines (336 loc) · 12.8 KB
/
mnist_mlp_urn.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
# coding: utf-8
# In[1]:
# get_ipython().magic(u'matplotlib inline')
import mxnet as mx
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.cm as cm
# In[2]:
dev = mx.cpu()
batch_size = 100
data_shape = (1, 28, 28)
train_iter = mx.io.MNISTIter(
image = "../softwares/Thesis/adversarial/LearningwithaStrongAdversary-master/datasets/mnist/train-images.idx3-ubyte",
label = "../softwares/Thesis/adversarial/LearningwithaStrongAdversary-master/datasets/mnist/train-labels.idx1-ubyte",
input_shape = data_shape,
batch_size = batch_size,
shuffle = True,
flat = False)
val_iter = mx.io.MNISTIter(
image = "../softwares/Thesis/adversarial/LearningwithaStrongAdversary-master/datasets/mnist/t10k-images.idx3-ubyte",
label = "../softwares/Thesis/adversarial/LearningwithaStrongAdversary-master/datasets/mnist/t10k-labels.idx1-ubyte",
input_shape = data_shape,
batch_size = batch_size,
flat = False)
# In[3]:
def Softmax(theta):
max_val = np.max(theta, axis=1, keepdims=True)
tmp = theta - max_val
exp = np.exp(tmp)
norm = np.sum(exp, axis=1, keepdims=True)
return exp / norm
def LogLossGrad(alpha, label):
grad = np.copy(alpha)
for i in range(alpha.shape[0]):
grad[i, label[i]] -= 1.
return grad
def SGD(weight, grad, lr=0.1, grad_norm=batch_size):
weight[:] -= lr * grad / batch_size
def CalAcc(pred_prob, label):
pred = np.argmax(pred_prob, axis=1)
return np.sum(pred == label) * 1.0
def CalLoss(pred_prob, label):
loss = 0.
for i in range(pred_prob.shape[0]):
loss += -np.log(max(pred_prob[i, label[i]], 1e-10))
return loss
# In[4]:
def acc_normal(model, val_iter, arg_map, grad_map):
val_iter.reset()
val_acc = 0.0
num_samp = 0
for dbatch in val_iter:
data = dbatch.data[0]
label = dbatch.label[0]
batch_size = label.asnumpy().shape[0]
arg_map["data"][:] = data
model.forward(is_train=False)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
val_acc += CalAcc(alpha, label.asnumpy())
num_samp += batch_size
return(val_acc / num_samp)
def acc_perb_L0(model, val_iter, coe_pb,arg_map, grad_map):
val_iter.reset()
val_acc = 0.0
num_samp = 0
nn=0
for dbatch in val_iter:
data = dbatch.data[0]
label = dbatch.label[0]
batch_size = label.asnumpy().shape[0]
arg_map["data"][:] = data
model.forward(is_train=True)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
grad = LogLossGrad(alpha, label.asnumpy())
out_grad[:] = grad
model.backward([out_grad])
noise = np.sign(grad_map["data"].asnumpy())
for j in range(batch_size):
if np.linalg.norm(noise[j].flatten(),2) ==0:
nn+=1
y = label.asnumpy()[j]
if (y == np.argmax(alpha[j])):
noise[j] = noise[j]/np.linalg.norm(noise[j].flatten(),2)
else:
noise[j] = 0
pdata = data.asnumpy() + coe_pb * noise
arg_map["data"][:] = pdata
model.forward(is_train=False)
raw_output = model.outputs[0].asnumpy()
pred = Softmax(raw_output)
val_acc += CalAcc(pred, label.asnumpy())
num_samp += batch_size
if nn>0:
print('L0 gradien being 0 :', nn)
return(val_acc / num_samp)
def acc_perb_L2(model, val_iter, coe_pb, arg_map, grad_map):
val_iter.reset()
val_acc = 0.0
num_batch = 0
nn=0
for dbatch in val_iter:
data = dbatch.data[0]
label = dbatch.label[0]
batch_size = label.asnumpy().shape[0]
arg_map["data"][:] = data
model.forward(is_train=True)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
grad = LogLossGrad(alpha, label.asnumpy())
out_grad[:] = grad
model.backward([out_grad])
noise = grad_map["data"].asnumpy()
for j in range(batch_size):
if np.linalg.norm(noise[j].flatten(),2) ==0:
nn+=1
y = label.asnumpy()[j]
if (y == np.argmax(alpha[j])):
noise[j] = noise[j]/np.linalg.norm(noise[j].flatten(),2)
else:
noise[j] = 0
pdata = data.asnumpy() + coe_pb * noise
arg_map["data"][:] = pdata
model.forward(is_train=False)
raw_output = model.outputs[0].asnumpy()
pred = Softmax(raw_output)
val_acc += CalAcc(pred, label.asnumpy()) / batch_size
num_batch += 1
if nn>0:
print('L2 gradien being 0 :', nn)
return(val_acc / num_batch)
def acc_perb_ks(model,n, val_iter, coe_pb, arg_map, grad_map):
val_iter.reset()
val_acc = 0.0
num_batch = 0
nn=0
for dbatch in val_iter:
data = dbatch.data[0]
label = dbatch.label[0]
batch_size = label.asnumpy().shape[0]
arg_map["data"][:] = data
model.forward(is_train=True)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
grad = LogLossGrad(alpha, label.asnumpy())
out_grad[:] = grad
model.backward([out_grad])
new_grad= grad_map["data"].asnumpy()
gresh=np.zeros((len(new_grad),784),dtype=float)
for i in np.arange(len(new_grad)):
temp_row=np.squeeze(new_grad[i])
gresh[i]=temp_row.flatten()
#n=int(np.floor(.95*784))
asorted=np.argsort(np.absolute(gresh),axis=-1, kind='quicksort')
mask=np.zeros((gresh.shape),dtype=float)
for i in np.arange(len(asorted)):
c=(asorted[i][-n:]).tolist()
for ind,val in enumerate(c):
mask[i,val]=1.0
temp_noise=np.multiply(gresh,mask)
noise=np.reshape(temp_noise, (new_grad.shape))
for j in range(batch_size):
if np.linalg.norm(noise[j].flatten(),2) ==0:
nn+=1
y = label.asnumpy()[j]
if (y == np.argmax(alpha[j])):
noise[j] = noise[j]/np.linalg.norm(noise[j].flatten(),2)
else:
noise[j] = 0
pdata = data.asnumpy() + coe_pb * noise
arg_map["data"][:] = pdata
model.forward(is_train=False)
raw_output = model.outputs[0].asnumpy()
pred = Softmax(raw_output)
val_acc += CalAcc(pred, label.asnumpy()) / batch_size
num_batch += 1
if nn>0:
print('ks gradien being 0 :', nn)
return(val_acc / num_batch)
def acc_perb_urn(model, val_iter, coe_pb, arg_map, grad_map):
val_iter.reset()
val_acc = 0.0
num_batch = 0
nn=0
for dbatch in val_iter:
data = dbatch.data[0]
label = dbatch.label[0]
batch_size = label.asnumpy().shape[0]
arg_map["data"][:] = data
model.forward(is_train=True)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
grad = LogLossGrad(alpha, label.asnumpy())
out_grad[:] = grad
model.backward([out_grad])
grad = grad_map["data"].asnumpy()
noise=np.random.randn(grad.shape[0],grad.shape[1],grad.shape[2],grad.shape[3])
pdata = data.asnumpy() + coe_pb * noise
arg_map["data"][:] = pdata
model.forward(is_train=False)
raw_output = model.outputs[0].asnumpy()
pred = Softmax(raw_output)
val_acc += CalAcc(pred, label.asnumpy()) / batch_size
num_batch += 1
return(val_acc / num_batch)
def acc_perb_alpha(model, val_iter, coe_pb,arg_map, grad_map):
val_iter.reset()
val_acc = 0.0
num_samp = 0
nn=0
for dbatch in val_iter:
data = dbatch.data[0]
label = dbatch.label[0]
batch_size = label.asnumpy().shape[0]
arg_map["data"][:] = data
T = np.zeros((10, batch_size, data_shape[1], data_shape[2], data_shape[3]))
noise = np.zeros(data.shape)
#===================
for i in range(10):
arg_map["data"][:] = data
model.forward(is_train=True)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
grad = LogLossGrad(alpha, i*np.ones(alpha.shape[0]))
for j in range(batch_size):
grad[j] = -alpha[j][i]*grad[j]
out_grad[:] = grad
model.backward([out_grad])
T[i] = grad_map["data"].asnumpy()
for j in range(batch_size):
y = label.asnumpy()[j]
if (y == np.argmax(alpha[j])):
perb_scale = np.zeros(10)
for i in range(10):
if (i == y):
perb_scale[i] = np.inf
else:
perb_scale[i] = (alpha[j][y] - alpha[j][i])/np.linalg.norm((T[i][j]-T[y][j]).flatten(),2)
noise[j] = T[np.argmin(perb_scale)][j]-T[y][j]
#====================
for j in range(batch_size):
if np.linalg.norm(noise[j].flatten(),2) ==0:
nn+=1
else:
noise[j] = noise[j]/np.linalg.norm(noise[j].flatten(),2)
pdata = data.asnumpy() + coe_pb * noise
arg_map["data"][:] = pdata
model.forward(is_train=False)
raw_output = model.outputs[0].asnumpy()
pred = Softmax(raw_output)
val_acc += CalAcc(pred, label.asnumpy()) /batch_size
num_samp += 1
if nn>0:
print('Alpha gradien being 0 :', nn)
return(val_acc / num_samp)
# # Ksup
# In[ ]:
# input
data = mx.symbol.Variable('data')
# first fullc
flatten = mx.symbol.Flatten(data=data)
fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=400)
relu1 = mx.symbol.Activation(data=fc1, act_type="relu")
fc2 = mx.symbol.FullyConnected(data=relu1, num_hidden=400)
relu2 = mx.symbol.Activation(data=fc2, act_type="relu")
# second fullc
fc3 = mx.symbol.FullyConnected(data=relu2, num_hidden=10)
# In[ ]:
data_shape = (batch_size, 1, 28, 28)
arg_names = fc3.list_arguments() # 'data'
arg_shapes, output_shapes, aux_shapes = fc3.infer_shape(data=data_shape)
arg_arrays = [mx.nd.zeros(shape, ctx=dev) for shape in arg_shapes]
grad_arrays = [mx.nd.zeros(shape, ctx=dev) for shape in arg_shapes]
reqs = ["write" for name in arg_names]
model = fc3.bind(ctx=dev, args=arg_arrays, args_grad = grad_arrays, grad_req=reqs)
arg_map = dict(zip(arg_names, arg_arrays))
grad_map = dict(zip(arg_names, grad_arrays))
data_grad = grad_map["data"]
out_grad = mx.nd.zeros(model.outputs[0].shape, ctx=dev)
# In[ ]:
for name in arg_names:
if "weight" in name:
arr = arg_map[name]
arr[:] = mx.rnd.uniform(-0.07, 0.07, arr.shape)
# In[ ]:
num_round = 40
train_acc = 0.
nbatch = 0
coe_pb = 1.7
lr = 0.1
for i in range(num_round):
train_loss = 0.
train_acc = 0.
nbatch = 0
train_iter.reset()
for dbatch in train_iter:
data = dbatch.data[0]
label = dbatch.label[0]
arg_map["data"][:] = data
model.forward(is_train=True)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
train_acc += CalAcc(alpha, label.asnumpy()) / batch_size
train_loss += CalLoss(alpha, label.asnumpy()) / batch_size
losGrad_theta = LogLossGrad(alpha, label.asnumpy())
out_grad[:] = losGrad_theta
model.backward([out_grad])
new_grad= data_grad.asnumpy()
noise=np.random.randn(new_grad.shape[0],new_grad.shape[1],new_grad.shape[2],new_grad.shape[3])
pdata = data.asnumpy() + coe_pb * noise
arg_map["data"][:] = pdata
model.forward(is_train=True)
theta = model.outputs[0].asnumpy()
alpha = Softmax(theta)
losGrad_theta = LogLossGrad(alpha, label.asnumpy())
out_grad[:] = losGrad_theta
model.backward([out_grad])
for name in arg_names:
if name != "data":
SGD(arg_map[name], grad_map[name], lr)
nbatch += 1
train_acc /= nbatch
train_loss /= nbatch
val_acc = acc_normal(model, val_iter,arg_map, grad_map)
print("Train Accuracy: %.4f\t Val Accuracy: %.4f\t Train Loss: %.5f" % (train_acc, val_acc, train_loss))
# In[ ]:
for i in range(10,90,10):
n=int(np.floor(i*784/100))
print('KS perturbation: %.4f' % acc_perb_ks(model,n, val_iter, 1.5,arg_map, grad_map))
print('Normal Validation: %.3f' % acc_normal(model,val_iter,arg_map, grad_map))
#print('Fixed set perturbation: %.3f' % acc_normal(model, perb_iter,arg_map, grad_map))
print('L0 perturbation: %.3f' % acc_perb_L0(model, val_iter, 1.5,arg_map, grad_map))
print('L2 perturbation: %.3f' % acc_perb_L2(model, val_iter, 1.5,arg_map, grad_map))
print('Alpha perturbation: %.3f' % acc_perb_alpha(model, val_iter, 1.5,arg_map, grad_map))
print('urn perturbation: %.3f' % acc_perb_urn(model, val_iter, coe_pb, arg_map, grad_map))