-
Notifications
You must be signed in to change notification settings - Fork 0
/
mlp_mnist.py
328 lines (256 loc) · 10.6 KB
/
mlp_mnist.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
import torch
import torchvision
from torchvision import datasets, transforms
import torch.nn as nn
import torch.nn.init as init
from torch.autograd import Variable
import torch.utils.data as Data
from mlp_class import MLP
import matplotlib.pyplot as plt
from utils import *
import numpy as np
torch.manual_seed(4)
EPOCH = 0
BATCH_SIZE = 128
LR = 0.1
DOWNLOAD_MNIST = True
save_model = 0
load_model = 1
en_plot = 0
SI = 1 # reminder for sign bit
IL = 3 # IL = torch.IntTensor([4]) Not including the signed bit
FL = 3 # FL = torch.IntTensor([12])
nonideal_train = 0 #fixed point train
nonideal_inference = 1 #fixed point inference
#enabling ntv which increases bit flip probability
en_ntv_inference = 0
en_ntv_train = 0
p_flip = 0.001
flip_len = IL + FL - 1#IL + FL + SI # flip all bits IL + FL + SI
grads = [] #gradients are inside
#train_data = torchvision.datasets.MNIST(
# root = './mnist/',
# train=True,
# transform=torchvision.transforms.ToTensor(),
# download=DOWNLOAD_MNIST,
#)
train_data = torchvision.datasets.MNIST(
root = './mnist/',
train=True,
transform=transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]),
download=DOWNLOAD_MNIST,
)
test_data = torchvision.datasets.MNIST(root='./mnist/', train=False,transform=transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]))
train_loader = Data.DataLoader(dataset=train_data, batch_size=BATCH_SIZE, shuffle=True)
test_x = Variable(torch.unsqueeze(test_data.test_data, dim=1),volatile=True).type(torch.FloatTensor)
test_y = test_data.test_labels
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('./mnist/', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=BATCH_SIZE, shuffle=False)
# defining hooks
def fixed_point_hook(self, input, output):
'''
if step % 50 == 0 and 'ReLU' == self.__class__.__name__:
print('sum > 1 :', ((output.data)>1).sum())
print('sum > 2 :', ((output.data) > 2).sum())
print('sum > 3 :', ((output.data) > 3).sum())
print('sum > 3.9 :', ((output.data) > 3.9).sum())
'''
# print('Inside ' + self.__class__.__name__ + ' forward')
# print('Inside ' + self.__class__.__name__ + ' backward')
apply_format_inplace('FXP', output.data, IL, FL)
def fixed_point_back_hook(self, grad_in, grad_out):
#print('Inside ' + self.__class__.__name__ + ' backward')
'''
print('grad_output size:', grad_out[0].size())
print('grad_input size:', grad_in[0].size())
print('grad_input size:', grad_in[1].size())
print('grad_input size:', grad_in[2].size())
print('grad_output size:', grad_out[0].size())
print(grad_out[0])
print('grad_input size:', grad_in[0].size())
print(grad_in[0])
print('grad_input size:', grad_in[1].size())
print('grad_input size:', grad_in[2].size())
print('grad_input tuple size: ', grad_in.__len__())
'''
#grads[self.__class__.__name__] = grad_in[0].data;
grads.append(grad_in[0].data);
if grad_in.__len__() > 1:
#print('grad_input size:', grad_in[1].size())
apply_format_inplace('FXP', grad_in[0].data, IL, FL)
apply_format_inplace('FXP', grad_in[1].data, IL, FL)#weight
apply_format_inplace('FXP', grad_in[2].data, IL, FL)#bias
#print(type(grad_in[0]))
#print(type(grad_in[1]))
#print(type(grad_in[2]))
return grad_in
def mse_loss(x, y, size_average=True, per_element=True):
#print(type(y))
#print(y.data.size())
#print(y.data.type())
#print(y.data.size())
y.data = y.data.unsqueeze(1)
y_onehot = torch.FloatTensor(BATCH_S, 10)
y_onehot.zero_()
y_onehot.scatter_(1, y.data, 1)
label = y_onehot
label = Variable(label)
loss_per_element = (x- label) ** 2
if per_element:
return loss_per_element.mean(1,keepdim=True)
if size_average:
return loss_per_element.mean()
return loss_per_element.sum()
mlp=MLP()
optimizer = torch.optim.SGD(mlp.parameters(), lr=LR)#, momentum=0.9)
loss_func = nn.CrossEntropyLoss()
#add backward hooker on each layer, to regulate the gradient to fixed point representation
if(nonideal_train):
mlp.hidden1.register_backward_hook(fixed_point_back_hook)
mlp.activation.register_backward_hook(fixed_point_back_hook)
mlp.hidden2.register_backward_hook(fixed_point_back_hook)
mlp.activation2.register_backward_hook(fixed_point_back_hook)
mlp.out.register_backward_hook(fixed_point_back_hook)
#add foward hooker on each layer, , to regulate the layer out to fixed point representation
mlp.hidden1.register_forward_hook(fixed_point_hook)
mlp.activation.register_forward_hook(fixed_point_hook)
mlp.hidden2.register_forward_hook(fixed_point_hook)
mlp.activation2.register_forward_hook(fixed_point_hook)
mlp.out.register_forward_hook(fixed_point_hook)
#>>>>>>>>Dump neural network weight to list
linear = list(mlp.hidden1.parameters())
init.constant(linear[1].data,0)
linear2 = list(mlp.hidden2.parameters())
init.constant(linear2[1].data,0)
out = list(mlp.out.parameters())
init.constant(out[1].data,0)
init.normal(linear[0].data, mean=0, std=0.01)
init.normal(linear2[0].data, mean=0, std=0.01)
init.normal(out[0].data, mean=0, std=0.01)
train_loss_y = np.array([])
train_loss_x = np.array([])
i=0
for epoch in range(EPOCH):
avg_loss = np.array([])
for step, (x,y) in enumerate(train_loader):
b_x = Variable(x,requires_grad = True)
b_y = Variable(y)
BATCH_S = b_x.size()[0]
#print(BATCH_S)
output = mlp(b_x)
#loss = loss_func(output, b_y)
loss = mse_loss(output,b_y)
#print(loss.size())
mlp.zero_grad()
'''
for i in range(BATCH_SIZE):
output[i:i+1].backward(torch.ones(2,10),retain_variables=True)
loss = loss_func(output,b_y)
mlp.zero_grad()
'''
ind = torch.zeros(b_x.size()[0],1)
ind[:,:] = 1
optimizer.zero_grad()
loss.backward(ind)
optimizer.step()
if nonideal_train:
apply_format_inplace('FXP', linear[0].data, IL, FL)#weight
apply_format_inplace('FXP', linear[1].data, IL, FL)#bias
apply_format_inplace('FXP', linear2[0].data, IL, FL)#weight
apply_format_inplace('FXP', linear2[1].data, IL, FL)#bias
apply_format_inplace('FXP', out[0].data, IL, FL)#weight
apply_format_inplace('FXP', out[1].data, IL, FL)#bias
if en_ntv_train:
apply_bitflip(linear[0].data, p_flip, IL, FL, flip_len)
apply_bitflip(linear2[0].data, p_flip, IL, FL, flip_len)
apply_bitflip(out[0].data, p_flip, IL, FL, flip_len)
if step % 50 == 0:
print('Epoch: ', epoch, '| train loss : %.8f' % loss.data[0][0])
avg_loss = np.append(avg_loss, loss.data[0][0])
'''
if step % 120 == 0:
print('printing!!!')
train_loss_y = np.append(train_loss_y,loss.data[0])
train_loss_x = np.append(train_loss_x,i)
plt.plot(train_loss_x, train_loss_y, 'r-', lw=2)
plt.pause(0.05)
#plt.show(block=False)
i = i + 1
train_loss_x = train_loss_x.astype(int)
'''
'''
if step % 50 == 0:
test_output = mlp(test_x)
pred_y = torch.max(test_output,1)[1].data.squeeze()
accuracy = sum(pred_y == test_y)/float(test_y.size(0))
print('Epoch: ',epoch,'| train loss : %.4f' %loss.data[0], '| test accuracy: %.2f' % accuracy)
'''
#per epoch updating......
if en_plot:
train_loss_y = np.append(train_loss_y,np.mean(avg_loss))
train_loss_x = np.append(train_loss_x, epoch)
plt.semilogy(train_loss_x, train_loss_y, 'r-', lw=2)
plt.xlabel('Epoch')
plt.ylabel('Training error')
plt.title('IL = 4, FL = 14')
plt.grid(True)
plt.pause(0.05)
plt.show()
if save_model:
print('Saving model......')
torch.save(mlp, 'mlp_backup2.pkl')
#>>>>>>>>Test neural network performance
if load_model:
print('Loading model......')
mlp = torch.load('mlp_backup2.pkl')
#>>>>>>>>>>>>>>>>>>>>>>>>Inference
linear = list(mlp.hidden1.parameters())
linear2 = list(mlp.hidden2.parameters())
out = list(mlp.out.parameters())
#>>>>>>>>Regulate trained neural network weight into fixed point counterpart, then do inference
if nonideal_inference:
apply_format_inplace('FXP',linear[0].data,IL,FL) # weight regulation
apply_format_inplace('FXP', linear[1].data, IL, FL) # bias regulation
apply_format_inplace('FXP',linear2[0].data,IL,FL) # weight regulation
apply_format_inplace('FXP', linear2[1].data, IL, FL) # bias regulation
apply_format_inplace('FXP',out[0].data,IL,FL) # weight regulation
apply_format_inplace('FXP', out[1].data, IL, FL) # bias regulation
apply_format_inplace('FXP',test_x.data,IL,FL) # input regulation
#apply_bitflip(linear[0].data,0.5,3,10,13)
if en_ntv_inference:
print("injecting bit flip fault model...")
apply_bitflip(linear[0].data,p_flip,IL,FL,flip_len)
apply_bitflip(linear2[0].data, p_flip, IL, FL, flip_len)
apply_bitflip(out[0].data, p_flip, IL, FL, flip_len)
#apply_bitflip(x,p,IL,FL,flip_length):
#print(type(test_x)) #<class 'torch.autograd.variable.Variable'>
#print(type(test_x.data)) #<class 'torch.FloatTensor'>
#>>>>>>>>add forward hooker on each layer if non-ideal inference
if nonideal_inference and (not nonideal_train ):
#print('applied!!!')
mlp.hidden1.register_forward_hook(fixed_point_hook)
mlp.activation.register_forward_hook(fixed_point_hook)
mlp.hidden2.register_forward_hook(fixed_point_hook)
mlp.activation2.register_forward_hook(fixed_point_hook)
mlp.out.register_forward_hook(fixed_point_hook)
#>>>>>>>> Testing!!!
#test_output = mlp(test_x)
#pred_y = torch.max(test_output, 1)[1].data.squeeze()
#accuracy = sum(pred_y == test_y)/float(test_y.size(0))
#print(accuracy, 'prediction accuracy!')
test_loss = 0
correct = 0
for data, target in test_loader:
#data, target = data.cuda(), target.cuda()
data, target = Variable(data, volatile=True), Variable(target)
output = mlp(data)
#test_loss += F.nll_loss(output, target, size_average=False).data[0] # sum up batch loss
pred = output.data.max(1, keepdim=True)[1] # get the index of the max log-probability
correct += pred.eq(target.data.view_as(pred)).cpu().sum()
#test_loss /= len(test_loader.dataset)
print('\nTest set: Accuracy: (%.4f)\n', 100. * correct / len(test_loader.dataset))
print('End of testing!')