/
toy.py
351 lines (277 loc) · 11.8 KB
/
toy.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
import torch
import torch.nn as nn
import numpy as np
import torchvision.models as models
from torch.autograd import Variable
from models.generator import TOYGenerator
import matplotlib.pyplot as plt
from torch.utils import data
import tqdm
from scipy.stats import truncnorm
import torch.optim as optim
import random
##############################################################
# Toy sample configuration
##############################################################
# the training setting
test_data_size = 500
test_batch_size = 8
test_epoch = 200
test_iter = 50
def own_loss(A, B):
"""
L-2 loss between A and B normalized by length.
Shape of A should be (features_num, ), shape of B should be (batch_size, features_num)
"""
return (A - B).norm()**2 / B.size(0)
class output_hook(object):
"""
Forward_hook used to get the output of the intermediate layer.
"""
def __init__(self):
super(output_hook, self).__init__()
self.outputs = None
def hook(self, module, input, output):
self.outputs = output
def clear(self):
self.outputs = None
##############################################################
# Real Data to train FP model
##############################################################
# generate 500 data
noisex = np.random.uniform(-4, 4, test_data_size)
noisey = np.random.uniform(-4, 4, test_data_size)
GTdata = list(zip(noisex, noisey))
label = ((noisex * noisey) > 0) * 1
plt.figure()
plt.subplot(2, 2, 1)
plt.title("real_data")
plt.scatter(noisex, noisey, s=50, c=label, alpha=.5)
##############################################################
# FP Model Part
##############################################################
# train fp model first
FP_Model = nn.Sequential(
nn.Linear(2, 128),
nn.BatchNorm1d(128, 0.8),
nn.ReLU(inplace=True),
nn.Linear(128, 256),
nn.BatchNorm1d(256, 0.8),
nn.ReLU(inplace=True),
nn.Linear(256, 512),
nn.BatchNorm1d(512, 0.8),
nn.ReLU(inplace=True),
nn.Linear(512, 2),
)
FloatTensor = torch.cuda.FloatTensor
LongTensor = torch.cuda.LongTensor
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(FP_Model.parameters(), lr = 0.001, weight_decay=0.0001, momentum = 0.9)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=(test_epoch//3))
# train FP model first
toy_dataset = data.TensorDataset(FloatTensor(GTdata), FloatTensor(label))
toy_dataloader = data.DataLoader(toy_dataset, batch_size = 10, shuffle = True)
FP_Model.cuda()
FP_Model.train()
print("Train FP model")
for _ in tqdm.tqdm(range(test_epoch), position = 0, leave=True):
for (input_data, input_label) in toy_dataloader:
input_data = input_data.cuda()
input_label = input_label.cuda()
pred = FP_Model(input_data)
loss = criterion(pred, input_label.long())
optimizer.zero_grad()
loss.backward()
optimizer.step()
##############################################################
# Gaussian Test data Part
##############################################################
# generate noise data
test_gaussian_noise = np.random.normal(0, 1, (test_data_size, 2))
test_gaussian_label = np.random.uniform(0, 1, test_data_size) > 0.5
plt.subplot(2, 2, 2)
print(test_gaussian_noise.shape)
print(test_gaussian_label.shape)
plt.scatter(test_gaussian_noise[:, 0], test_gaussian_noise[:, 1], s=50, c=test_gaussian_label, alpha=.5)
plt.xlim(-4,4)
plt.ylim(-4,4)
plt.title("gaussian")
##############################################################
# ZeroQ Part
##############################################################
# use copy to prevent modification (Normally it will not be modified)
zeroQ_gaussian_input_data = list(test_gaussian_noise).copy()
zeroQ_gaussian_input_label = test_gaussian_label.copy()
random.shuffle(zeroQ_gaussian_input_data)
toy_dataset = data.TensorDataset(FloatTensor(zeroQ_gaussian_input_data), LongTensor(test_gaussian_label))
toy_dataloader = data.DataLoader(toy_dataset, batch_size = test_batch_size, shuffle = True)
print("ZeroQ training...")
# zero Q does not need input_label
# reference from https://github.com/amirgholami/ZeroQ/blob/f91092e0dfee676f4f667e04b2228543b27f9ce1/distill_data.py#L53
FP_Model = FP_Model.eval()
hooks, hook_handles, bn_stats = [], [], []
# get number of BatchNorm layers in the model
layers = sum([
1 if isinstance(layer, nn.BatchNorm1d) else 0
for layer in FP_Model.modules()
])
eps = 0.8
for n, m in FP_Model.named_modules():
if isinstance(m, nn.Linear) and len(hook_handles) < layers:
hook = output_hook()
hooks.append(hook)
hook_handles.append(m.register_forward_hook(hook.hook))
if isinstance(m, nn.BatchNorm1d):
# get the statistics in the BatchNorm layers
bn_stats.append(
(m.running_mean.detach().clone().flatten().cuda(),
torch.sqrt(m.running_var + eps).detach().clone().flatten().cuda()))
assert len(hooks) == len(bn_stats)
zeroQ_gaussian_data = []
zeroQ_gaussian_label = []
for idx, (input_data, input_label) in enumerate(toy_dataloader):
input_data = input_data.cuda()
input_data.requires_grad = True
FP_Model.zero_grad()
crit = nn.CrossEntropyLoss().cuda()
# optimizer = torch.optim.SGD([input_data], lr=0.001, momentum=0.9)
# scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=(test_epoch//3))
optimizer = optim.Adam([input_data], lr=0.005)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
min_lr=1e-5,
verbose=False,
patience=100)
input_mean = torch.FloatTensor([0.0]).cuda()
input_std = torch.FloatTensor([1.0]).cuda()
# make total data num same as generator
# total is test_epoch * iter * batch
for _ in tqdm.tqdm(range(test_epoch*test_iter)):
FP_Model.zero_grad()
optimizer.zero_grad()
# clear hook
for hook in hooks:
hook.clear()
output = FP_Model(input_data)
mean_loss = 0
std_loss = 0
# compute the loss according to the BatchNorm statistics and the statistics of intermediate output
for cnt, (bn_stat, hook) in enumerate(zip(bn_stats, hooks)):
tmp_output = hook.outputs
bn_mean, bn_std = bn_stat[0], bn_stat[1]
# get batch's norm
tmp_mean = torch.mean(tmp_output , dim = 0)
tmp_std = torch.sqrt(torch.var(tmp_output, dim = 0) + eps)
mean_loss += own_loss(bn_mean, tmp_mean)
std_loss += own_loss(bn_std, tmp_std)
tmp_mean = torch.mean(input_data, dim = -1)
tmp_std = torch.sqrt(torch.var(input_data, dim = -1) + eps)
mean_loss += own_loss(input_mean, tmp_mean)
std_loss += own_loss(input_std, tmp_std)
total_loss = mean_loss + std_loss
# update the distilled data
total_loss.backward()
optimizer.step()
scheduler.step(total_loss.item())
zeroQ_gaussian_data.extend(input_data.detach().clone().tolist())
zeroQ_gaussian_label.extend(input_label.detach().clone().tolist())
for handle in hook_handles:
handle.remove()
zeroQ_x, zeroQ_y = list(zip(*zeroQ_gaussian_data))
plt.subplot(2, 2, 3)
plt.scatter(list(zeroQ_x), list(zeroQ_y), s = 50, c = zeroQ_gaussian_label, alpha = .5)
plt.xlim(-4, 4)
plt.ylim(-4, 4)
plt.title("zeroQ")
##############################################################
# Generative Fake Data Part
##############################################################
print("Train Generative data.....")
im_shape = (2)
toy_g = TOYGenerator(n_classes = 2, in_channel = 2, img_shape = im_shape)
toy_g.train()
toy_g = toy_g.cuda()
criterion = nn.CrossEntropyLoss()
##############################################################
# Generator training
##############################################################
print("Train Generator")
optimizer = torch.optim.Adam(toy_g.parameters(), lr=1e-3)
# down 3 times
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=(100))
FP_Model = FP_Model.eval()
hooks, hook_handles, bn_stats = [], [], []
# get number of BatchNorm layers in the model
layers = sum([
1 if isinstance(layer, nn.BatchNorm1d) else 0
for layer in FP_Model.modules()
])
eps = 0.8
for n, m in FP_Model.named_modules():
if isinstance(m, nn.Linear) and len(hook_handles) < layers:
hook = output_hook()
hooks.append(hook)
hook_handles.append(m.register_forward_hook(hook.hook))
if isinstance(m, nn.BatchNorm1d):
# get the statistics in the BatchNorm layers
bn_stats.append(
(m.running_mean.detach().clone().flatten().cuda(),
torch.sqrt(m.running_var + eps).detach().clone().flatten().cuda()))
assert len(hooks) == len(bn_stats)
# clear hook
for hook in hooks:
hook.clear()
criterion = nn.CrossEntropyLoss()
# set same iteration as zeroQ
for epoch in range(test_epoch):
# for _ in tqdm.tqdm(range(test_epoch * (test_data_size//test_batch_size))):
pbar = tqdm.trange(test_iter)
for _ in pbar:
input_mean = torch.FloatTensor([0.0]).cuda()
input_std = torch.FloatTensor([1.0]).cuda()
FP_Model.zero_grad()
optimizer.zero_grad()
train_gaussian_noise = np.random.normal(0, 1, (test_batch_size, 2))
train_gaussian_label = np.random.randint(0, 2, test_batch_size)
input_data = Variable(FloatTensor(train_gaussian_noise)).cuda()
input_label = Variable(LongTensor(train_gaussian_label)).cuda()
fake_data = toy_g(input_data, input_label)
fake_label = FP_Model(fake_data)
mean_loss = 0
std_loss = 0
# compute the loss according to the BatchNorm statistics and the statistics of intermediate output
for cnt, (bn_stat, hook) in enumerate(zip(bn_stats, hooks)):
tmp_output = hook.outputs
bn_mean, bn_std = bn_stat[0], bn_stat[1]
# get batch's norm
tmp_mean = torch.mean(tmp_output , dim = 0)
tmp_std = torch.sqrt(torch.var(tmp_output, dim = 0) + eps)
mean_loss += own_loss(bn_mean, tmp_mean)
std_loss += own_loss(bn_std, tmp_std)
tmp_mean = torch.mean(input_data, dim = -1)
tmp_std = torch.sqrt(torch.var(input_data, dim = -1) + eps)
mean_loss += own_loss(input_mean, tmp_mean)
std_loss += own_loss(input_std, tmp_std)
g_loss = criterion(fake_label, input_label)
total_loss = g_loss + 0.1 * (mean_loss + std_loss)
total_loss.backward()
optimizer.step()
scheduler.step()
##############################################################
# Inference Generator with test_gaussian dataset
##############################################################
toy_g.eval()
test_inputs = FloatTensor(test_gaussian_noise)
test_labels = FloatTensor(test_gaussian_label)
fake_data_list = []
for idx, test_d in enumerate(test_inputs):
test_input = test_d.unsqueeze(0).cuda()
test_label = test_labels[idx].unsqueeze(0).cuda()
fake_data = toy_g(test_input, test_label.long())
fake_data_list.append(fake_data.tolist()[0])
fake_x, fake_y = zip(*fake_data_list)
plt.subplot(2, 2, 4)
plt.scatter(list(fake_x), list(fake_y), s = 50, c = test_gaussian_label, alpha = .5)
plt.xlim(-4, 4)
plt.ylim(-4, 4)
plt.title("Fake data")
plt.savefig('toy.png')