/
mnist.py
189 lines (145 loc) · 6.11 KB
/
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
# Copyright 2015 Matthieu Courbariaux, Zhouhan Lin
"""
This file is adapted from BinaryConnect:
https://github.com/MatthieuCourbariaux/BinaryConnect
Running this script should reproduce the results of a feed forward net trained
on MNIST.
To train a vanilla feed forward net with ordinary backprop:
1. type "git checkout fullresolution" to switch to the "fullresolution" branch
2. execute "python mnist.py"
To train a feed forward net with Binary Connect + quantized backprop:
1. type "git checkout binary" to switch to the "binary" branch
2. execute "python mnist.py"
To train a feed forward net with Ternary Connect + quantized backprop:
1. type "git checkout ternary" to switch to the "ternary" branch
2. execute "python mnist.py"
"""
import gzip
import cPickle
import numpy as np
import os
import os.path
import sys
import time
from trainer import Trainer
from model import Network
from layer import linear_layer, ReLU_layer
from pylearn2.datasets.mnist import MNIST
from pylearn2.utils import serial
if __name__ == "__main__":
rng = np.random.RandomState(1234)
train_set_size = 50000
# data augmentation
zero_pad = 0
affine_transform_a = 0
affine_transform_b = 0
horizontal_flip = False
# batch
# keep a multiple a factor of 10000 if possible
# 10000 = (2*5)^4
batch_size = 200
number_of_batches_on_gpu = train_set_size/batch_size
BN = True
BN_epsilon=1e-4 # for numerical stability
BN_fast_eval= True
dropout_input = 1.
dropout_hidden = 1.
shuffle_examples = True
shuffle_batches = False
# Termination criteria
n_epoch = 1000
monitor_step = 2
# LR
LR = .3
LR_fin = .01
LR_decay = (LR_fin/LR)**(1./n_epoch)
M= 0.
# architecture
n_inputs = 784
n_units = 1024
n_classes = 10
n_hidden_layer = 3
# BinaryConnect
BinaryConnect = True
stochastic = True
# Old hyperparameters
binary_training=False
stochastic_training=False
binary_test=False
stochastic_test=False
if BinaryConnect == True:
binary_training=True
if stochastic == True:
stochastic_training=True
else:
binary_test=True
print 'Loading the dataset'
train_set = MNIST(which_set= 'train', start=0, stop = train_set_size, center = True)
valid_set = MNIST(which_set= 'train', start=50000, stop = 60000, center = True)
test_set = MNIST(which_set= 'test', center = True)
# bc01 format
train_set.X = train_set.X.reshape(train_set_size,1,28,28)
valid_set.X = valid_set.X.reshape(10000,1,28,28)
test_set.X = test_set.X.reshape(10000,1,28,28)
# flatten targets
train_set.y = np.hstack(train_set.y)
valid_set.y = np.hstack(valid_set.y)
test_set.y = np.hstack(test_set.y)
# Onehot the targets
train_set.y = np.float32(np.eye(10)[train_set.y])
valid_set.y = np.float32(np.eye(10)[valid_set.y])
test_set.y = np.float32(np.eye(10)[test_set.y])
# for hinge loss
train_set.y = 2* train_set.y - 1.
valid_set.y = 2* valid_set.y - 1.
test_set.y = 2* test_set.y - 1.
print 'Creating the model'
class PI_MNIST_model(Network):
def __init__(self, rng):
Network.__init__(self, n_hidden_layer = n_hidden_layer, BN = BN)
print " Fully connected layer 1:"
self.layer.append(ReLU_layer(rng = rng, n_inputs = n_inputs, n_units = n_units,
BN = BN, BN_epsilon=BN_epsilon, dropout=dropout_input,
binary_training=binary_training, stochastic_training=stochastic_training,
binary_test=binary_test, stochastic_test=stochastic_test))
for k in range(n_hidden_layer-1):
print " Fully connected layer "+ str(k) +":"
self.layer.append(ReLU_layer(rng = rng, n_inputs = n_units, n_units = n_units,
BN = BN, BN_epsilon=BN_epsilon, dropout=dropout_hidden,
binary_training=binary_training, stochastic_training=stochastic_training,
binary_test=binary_test, stochastic_test=stochastic_test))
print " L2 SVM layer:"
self.layer.append(linear_layer(rng = rng, n_inputs = n_units, n_units = n_classes,
BN = BN, BN_epsilon=BN_epsilon, dropout=dropout_hidden,
binary_training=binary_training, stochastic_training=stochastic_training,
binary_test=binary_test, stochastic_test=stochastic_test))
model = PI_MNIST_model(rng = rng)
print 'Creating the trainer'
trainer = Trainer(rng = rng,
train_set = train_set, valid_set = valid_set, test_set = test_set,
model = model, load_path = None, save_path = None,
zero_pad=zero_pad,
affine_transform_a=affine_transform_a, # a is (more or less) the rotations
affine_transform_b=affine_transform_b, # b is the translations
horizontal_flip=horizontal_flip,
LR = LR, LR_decay = LR_decay, LR_fin = LR_fin,
M = M,
BN = BN, BN_fast_eval=BN_fast_eval,
batch_size = batch_size, number_of_batches_on_gpu = number_of_batches_on_gpu,
n_epoch = n_epoch, monitor_step = monitor_step,
shuffle_batches = shuffle_batches, shuffle_examples = shuffle_examples)
print 'Building'
trainer.build()
print 'Training'
start_time = time.clock()
trainer.train()
end_time = time.clock()
print 'The training took %i seconds'%(end_time - start_time)
print 'Display weights'
import matplotlib.pyplot as plt
import matplotlib.cm as cm
from filter_plot import tile_raster_images
W = np.transpose(model.layer[0].W.get_value())
W = tile_raster_images(W,(28,28),(4,4),(2, 2))
plt.imshow(W, cmap = cm.Greys_r)
plt.savefig(core_path + '_features.png')