-
Notifications
You must be signed in to change notification settings - Fork 1.3k
/
test.py
210 lines (182 loc) · 7.22 KB
/
test.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
"""
Testing code for different neural network configurations.
Adapted for Python 3.5.2
Usage in shell:
python3.5 test.py
Network (network.py and network2.py) parameters:
2nd param is epochs count
3rd param is batch size
4th param is learning rate (eta)
Author:
Michał Dobrzański, 2016
"""
# ----------------------
# - read the input data:
'''
import mnist_loader
training_data, validation_data, test_data = mnist_loader.load_data_wrapper()
training_data = list(training_data)
'''
# ---------------------
# - network.py example:
#import network
'''
net = network.Network([784, 30, 10])
net.SGD(training_data, 30, 10, 3.0, test_data=test_data)
'''
# ----------------------
# - network2.py example:
#import network2
'''
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
#net.large_weight_initializer()
net.SGD(training_data, 30, 10, 0.1, lmbda = 5.0,evaluation_data=validation_data,
monitor_evaluation_accuracy=True)
'''
# chapter 3 - Overfitting example - too many epochs of learning applied on small (1k samples) amount od data.
# Overfitting is treating noise as a signal.
'''
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
net.large_weight_initializer()
net.SGD(training_data[:1000], 400, 10, 0.5, evaluation_data=test_data,
monitor_evaluation_accuracy=True,
monitor_training_cost=True)
'''
# chapter 3 - Regularization (weight decay) example 1 (only 1000 of training data and 30 hidden neurons)
'''
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
net.large_weight_initializer()
net.SGD(training_data[:1000], 400, 10, 0.5,
evaluation_data=test_data,
lmbda = 0.1, # this is a regularization parameter
monitor_evaluation_cost=True,
monitor_evaluation_accuracy=True,
monitor_training_cost=True,
monitor_training_accuracy=True)
'''
# chapter 3 - Early stopping implemented
'''
net = network2.Network([784, 30, 10], cost=network2.CrossEntropyCost)
net.SGD(training_data[:1000], 30, 10, 0.5,
lmbda=5.0,
evaluation_data=validation_data,
monitor_evaluation_accuracy=True,
monitor_training_cost=True,
early_stopping_n=10)
'''
# chapter 4 - The vanishing gradient problem - deep networks are hard to train with simple SGD algorithm
# this network learns much slower than a shallow one.
'''
net = network2.Network([784, 30, 30, 30, 30, 10], cost=network2.CrossEntropyCost)
net.SGD(training_data, 30, 10, 0.1,
lmbda=5.0,
evaluation_data=validation_data,
monitor_evaluation_accuracy=True)
'''
# ----------------------
# Theano and CUDA
# ----------------------
"""
This deep network uses Theano with GPU acceleration support.
I am using Ubuntu 16.04 with CUDA 7.5.
Tutorial:
http://deeplearning.net/software/theano/install_ubuntu.html#install-ubuntu
The following command will update only Theano:
sudo pip install --upgrade --no-deps theano
The following command will update Theano and Numpy/Scipy (warning bellow):
sudo pip install --upgrade theano
"""
"""
Below, there is a testing function to check whether your computations have been made on CPU or GPU.
If the result is 'Used the cpu' and you want to have it in gpu, do the following:
1) install theano:
sudo python3.5 -m pip install Theano
2) download and install the latest cuda:
https://developer.nvidia.com/cuda-downloads
I had some issues with that, so I followed this idea (better option is to download the 1,1GB package as .run file):
http://askubuntu.com/questions/760242/how-can-i-force-16-04-to-add-a-repository-even-if-it-isnt-considered-secure-eno
You may also want to grab the proper NVidia driver, choose it form there:
System Settings > Software & Updates > Additional Drivers.
3) should work, run it with:
THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python3.5 test.py
http://deeplearning.net/software/theano/tutorial/using_gpu.html
4) Optionally, you can add cuDNN support from:
https://developer.nvidia.com/cudnn
"""
def testTheano():
from theano import function, config, shared, sandbox
import theano.tensor as T
import numpy
import time
print("Testing Theano library...")
vlen = 10 * 30 * 768 # 10 x #cores x # threads per core
iters = 1000
rng = numpy.random.RandomState(22)
x = shared(numpy.asarray(rng.rand(vlen), config.floatX))
f = function([], T.exp(x))
print(f.maker.fgraph.toposort())
t0 = time.time()
for i in range(iters):
r = f()
t1 = time.time()
print("Looping %d times took %f seconds" % (iters, t1 - t0))
print("Result is %s" % (r,))
if numpy.any([isinstance(x.op, T.Elemwise) for x in f.maker.fgraph.toposort()]):
print('Used the cpu')
else:
print('Used the gpu')
# Perform check:
#testTheano()
# ----------------------
# - network3.py example:
import network3
from network3 import Network, ConvPoolLayer, FullyConnectedLayer, SoftmaxLayer # softmax plus log-likelihood cost is more common in modern image classification networks.
# read data:
training_data, validation_data, test_data = network3.load_data_shared()
# mini-batch size:
mini_batch_size = 10
# chapter 6 - shallow architecture using just a single hidden layer, containing 100 hidden neurons.
'''
net = Network([
FullyConnectedLayer(n_in=784, n_out=100),
SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.1, validation_data, test_data)
'''
# chapter 6 - 5x5 local receptive fields, 20 feature maps, max-pooling layer 2x2
'''
net = Network([
ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),
filter_shape=(20, 1, 5, 5),
poolsize=(2, 2)),
FullyConnectedLayer(n_in=20*12*12, n_out=100),
SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.1, validation_data, test_data)
'''
# chapter 6 - inserting a second convolutional-pooling layer to the previous example => better accuracy
'''
net = Network([
ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),
filter_shape=(20, 1, 5, 5),
poolsize=(2, 2)),
ConvPoolLayer(image_shape=(mini_batch_size, 20, 12, 12),
filter_shape=(40, 20, 5, 5),
poolsize=(2, 2)),
FullyConnectedLayer(n_in=40*4*4, n_out=100),
SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.1, validation_data, test_data)
'''
# chapter 6 - rectified linear units and some l2 regularization (lmbda=0.1) => even better accuracy
from network3 import ReLU
net = Network([
ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),
filter_shape=(20, 1, 5, 5),
poolsize=(2, 2),
activation_fn=ReLU),
ConvPoolLayer(image_shape=(mini_batch_size, 20, 12, 12),
filter_shape=(40, 20, 5, 5),
poolsize=(2, 2),
activation_fn=ReLU),
FullyConnectedLayer(n_in=40*4*4, n_out=100, activation_fn=ReLU),
SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
net.SGD(training_data, 60, mini_batch_size, 0.03, validation_data, test_data, lmbda=0.1)