forked from benanne/nervana_theano
-
Notifications
You must be signed in to change notification settings - Fork 0
/
mnist_conv_nervana.py
146 lines (114 loc) · 4.03 KB
/
mnist_conv_nervana.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
from __future__ import print_function
import lasagne
import theano
import theano.tensor as T
import time
from mnist import _load_data
from mnist import create_iter_functions
from mnist import train
from lasagne.layers.cuda_convnet import bc01_to_c01b, c01b_to_bc01, MaxPool2DCCLayer
from nervana_theano import layers
NUM_EPOCHS = 500
BATCH_SIZE = 600
LEARNING_RATE = 0.01
MOMENTUM = 0.9
def load_data():
data = _load_data()
X_train, y_train = data[0]
X_valid, y_valid = data[1]
X_test, y_test = data[2]
# reshape for convolutions
X_train = X_train.reshape((X_train.shape[0], 1, 28, 28))
X_valid = X_valid.reshape((X_valid.shape[0], 1, 28, 28))
X_test = X_test.reshape((X_test.shape[0], 1, 28, 28))
return dict(
X_train=theano.shared(lasagne.utils.floatX(X_train)),
y_train=T.cast(theano.shared(y_train), 'int32'),
X_valid=theano.shared(lasagne.utils.floatX(X_valid)),
y_valid=T.cast(theano.shared(y_valid), 'int32'),
X_test=theano.shared(lasagne.utils.floatX(X_test)),
y_test=T.cast(theano.shared(y_test), 'int32'),
num_examples_train=X_train.shape[0],
num_examples_valid=X_valid.shape[0],
num_examples_test=X_test.shape[0],
input_height=X_train.shape[2],
input_width=X_train.shape[3],
output_dim=10,
)
def build_model(input_width, input_height, output_dim,
batch_size=BATCH_SIZE):
l_in = lasagne.layers.InputLayer(
shape=(batch_size, 1, input_width, input_height),
)
l_in_c01b = bc01_to_c01b(l_in)
l_conv1 = layers.NervanaConvLayer(
l_in_c01b,
num_filters=32,
filter_size=(5, 5),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform(),
dimshuffle=False,
)
l_pool1 = MaxPool2DCCLayer(l_conv1, ds=(2, 2), dimshuffle=False)
l_conv2 = layers.NervanaConvLayer(
l_pool1,
num_filters=32,
filter_size=(5, 5),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform(),
dimshuffle=False,
)
l_pool2 = MaxPool2DCCLayer(l_conv2, ds=(2, 2), dimshuffle=False)
l_pool2_bc01 = c01b_to_bc01(l_pool2)
l_hidden1 = lasagne.layers.DenseLayer(
l_pool2_bc01,
num_units=256,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform(),
)
l_hidden1_dropout = lasagne.layers.DropoutLayer(l_hidden1, p=0.5)
# l_hidden2 = lasagne.layers.DenseLayer(
# l_hidden1_dropout,
# num_units=256,
# nonlinearity=lasagne.nonlinearities.rectify,
# )
# l_hidden2_dropout = lasagne.layers.DropoutLayer(l_hidden2, p=0.5)
l_out = lasagne.layers.DenseLayer(
l_hidden1_dropout,
num_units=output_dim,
nonlinearity=lasagne.nonlinearities.softmax,
W=lasagne.init.GlorotUniform(),
)
return l_out
def main(num_epochs=NUM_EPOCHS):
print("Loading data...")
dataset = load_data()
print("Building model and compiling functions...")
output_layer = build_model(
input_height=dataset['input_height'],
input_width=dataset['input_width'],
output_dim=dataset['output_dim'],
)
iter_funcs = create_iter_functions(
dataset,
output_layer,
X_tensor_type=T.tensor4,
)
print("Starting training...")
now = time.time()
try:
for epoch in train(iter_funcs, dataset):
print("Epoch {} of {} took {:.3f}s".format(
epoch['number'], num_epochs, time.time() - now))
now = time.time()
print(" training loss:\t\t{:.6f}".format(epoch['train_loss']))
print(" validation loss:\t\t{:.6f}".format(epoch['valid_loss']))
print(" validation accuracy:\t\t{:.2f} %%".format(
epoch['valid_accuracy'] * 100))
if epoch['number'] >= num_epochs:
break
except KeyboardInterrupt:
pass
return output_layer
if __name__ == '__main__':
main()