-
Notifications
You must be signed in to change notification settings - Fork 6
/
vanillaNN.py
166 lines (130 loc) · 4.38 KB
/
vanillaNN.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
import theano
import theano.tensor as T
import numpy
import uuid
import time
from theano_tools import shared, HiddenLayer, StackModel, RandomStreams, momentum,\
GenericClassificationDataset, tools, gradient_descent, reinforce_no_baseline, \
InputSparseHiddenLayer, reinforce_no_baseline_momentum
import matplotlib
matplotlib.use('Agg')
from matplotlib import pyplot
import cPickle as pickle
# symbolic RNG
srng = RandomStreams(142857)
class Dropout:
def __init__(self):
pass
def __call__(self, x):
return (srng.uniform(x.shape) < 0.5) * x
def build_model(new_model=True):
momentum_epsilon = 0.9
nhidden = [64,64]
L2reg = 0.001
vanilla = True
hyperparams = locals()
if new_model:
expid = str(uuid.uuid4())
import os
import os.path
code = file(os.path.abspath(__file__),'r').read()
os.mkdir(expid)
os.chdir(expid)
file('code.py','w').write(code)
print expid
f = file("params.txt",'w')
for i in hyperparams:
f.write("%s:%s\n"%(i,str(hyperparams[i])))
f.close()
params = []
shared.bind(params)
rect = lambda x:T.maximum(0,x)
act = T.tanh
model = StackModel([HiddenLayer(32*32*3, nhidden[0], act),
Dropout(),
HiddenLayer(nhidden[0], nhidden[1], act),
Dropout(),
HiddenLayer(nhidden[-1], 10, T.nnet.softmax)])
x = T.matrix()
y = T.ivector()
lr = T.scalar()
y_hat, = model(x)
loss = T.nnet.categorical_crossentropy(y_hat, y)
cost = T.sum(loss)
l2 = lambda x:sum([T.sum(i**2) for i in x])
updates = []
error = T.sum(T.neq(y_hat.argmax(axis=1), y))
nn_regularization = L2reg * l2(params)
grads = T.grad(cost + nn_regularization, params)
updates += gradient_descent(params, grads, lr)
learn = theano.function([x,y,lr], [cost, error], updates=updates, allow_input_downcast=True)
test = theano.function([x,y], [cost, error], allow_input_downcast=True)
return model,learn,test
def main():
data = GenericClassificationDataset("cifar10", "cifar_10_shuffled.pkl")
N = data.train[0].shape[0] * 1.
model, learn, test = build_model()
experiment = {"results":None,
}
lr = 0.001 # * 100 / (i+100)
costs = []
errors = []
valid_costs = []
valid_errors = []
for i in range(1000):
epoch = i
cost = 0
error = 0
for x,y in data.trainMinibatches(128):
c,e = learn(x,y,lr)
cost += c
error += e
t0 = time.time()
valid_error, valid_cost = data.validate(test, 50)
valid_time = time.time() - t0
print
print i, cost/N, error/N
print valid_error, valid_cost, valid_time
errors.append(error/N)
costs.append(cost/N)
valid_errors.append(valid_error)
valid_costs.append(valid_cost)
tools.export_feature_image(model.layers[0].W, "W_img.png", (32,32,3))
tools.export_multi_plot1d([errors, valid_errors], "errors.png", "error")
tools.export_multi_plot1d([costs, valid_costs], "costs.png", "cost")
experiment["results"] = [valid_costs, valid_errors, costs, errors]
experiment["valid_time"] = valid_time
pickle.dump(experiment, file("experiment.pkl",'w'),-1)
shared.exportToFile("weights.pkl")
#video.stdin.close()
#video.wait()
def test(expid):
# to test:
# OMP_NUM_THREADS=1 THEANO_FLAGS=device=cpu taskset -c 0 python $(expip)/code.py $(expid)
import os
os.chdir(expid)
print "loading data"
data = GenericClassificationDataset("cifar10", "../cifar_10_shuffled.pkl")
print "building model"
model,learn,test = build_model(False)
print "importing weights"
shared.importFromFile("weights.pkl")
print "testing"
import time
t0 = time.time()
test_error, test_cost = data.doTest(test, 50)
t1 = time.time()
print "Error, cost, time(s)"
print test_error, test_cost, t1-t0
specialized_test_time = t1-t0
normal_test_time = t1-t0
f= file("test_results.txt",'w')
f.write("specialized:%f\ntheano:%f\nerror:%f\n"%(specialized_test_time, normal_test_time, test_error))
f.close()
if __name__ == "__main__":
import sys
print sys.argv
if len(sys.argv) <= 1:
main()
else:
test(sys.argv[1])