forked from Jashmi/FitNets
/
cnn_3.py
221 lines (158 loc) · 7.7 KB
/
cnn_3.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
import theano
from theano import tensor as T
from theano.tensor.nnet import conv2d
from theano.tensor.signal import downsample
from logreg import load_data, LogReg
import sys, numpy
from mlp import HiddenLayer
import timeit , pickle
sys.path.insert(0, "/home/mayank/Downloads/computer_vision/Learning/")
class LeNetConvPoolLayer(object):
def __init__(self,rng,input,filter_shape,image_shape,poolsize=(2,2),border_mode = 'valid'):
'''
image_shape - no. of images in a batch, no. of input features(layer m-1) ,
image height, image width
filter_shape - no. of filters or the no. of output features (layer m), no. of input features(layer m-1),
filter height, filter width
'''
'''
For the case of input image(1st stage) it ensures that if the Image is RGB
then the filter depth is also 3 and if the image is B/W the filter depth is 1
For further layers
'''
assert image_shape[1]==filter_shape[1]
self.input = input
'''
Input sise for each filter multiplication (of 2 matrices)
during convolution.
The no. of elements that would feed back the gradient to each one - so kind of like normalising
'''
fan_in = numpy.prod(filter_shape[1:])
fan_out = numpy.prod(filter_shape[0]*filter_shape[2]*filter_shape[3])//numpy.prod(poolsize)
'''
Intialising weight and bias
'''
W_bound = numpy.sqrt(6./(fan_in+fan_out))
self.W = theano.shared(numpy.asarray(rng.uniform(low = -W_bound, high = W_bound, size = filter_shape),dtype = theano.config.floatX)
, borrow =True)
self.b = theano.shared(numpy.zeros(shape = filter_shape[0], dtype = theano.config.floatX), borrow =True)
'''
convolving input and filters
If I just write cone2d(input,..) instead of specifying by name that also works
'''
conv_out = conv2d(input=input,filters = self.W, filter_shape = filter_shape, input_shape = image_shape,border_mode = border_mode)
# maxpooling
pooled_out = downsample.max_pool_2d(input = conv_out,ds = poolsize, ignore_border = True, padding = (0,0) )
self.output = T.tanh(pooled_out + self.b.dimshuffle('x',0,'x','x'))
self.params = [self.W, self.b]
self.input = input
def evaluate_lenet5(learning_rate = 0.1, n_epochs=200, dataset='mnist.pkl.gz',nkerns = [16,16,16], batch_size = 500):
rng = numpy.random.RandomState(32324)
datasets = load_data(dataset)
train_set_x,train_set_y = datasets[0]
valid_set_x,valid_set_y = datasets[1]
test_set_x,test_set_y = datasets[2]
n_train_batches = train_set_x.get_value(borrow=True).shape[0]//batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0]//batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0]//batch_size
index = T.lscalar() # index for each mini batch
train_epoch = T.lscalar()
x = T.matrix('x')
y = T.ivector('y')
# ------------------------------- Building Model ----------------------------------
print "...Building the model"
# output image size = (28-5+1+4)/2 = 14
layer_0_input = x.reshape((batch_size,1,28,28))
layer_0 = LeNetConvPoolLayer(rng,input = layer_0_input, image_shape=(batch_size,1,28,28),
filter_shape=(nkerns[0],1,5,5),poolsize=(2,2), border_mode = 2)
#output image size = (14-3+1)/2 = 6
layer_1 = LeNetConvPoolLayer(rng, input = layer_0.output, image_shape = (batch_size, nkerns[0],14,14),
filter_shape = (nkerns[1],nkerns[0],3,3), poolsize=(2,2) )
#output image size = (6-3+1)/2 = 2
layer_2 = LeNetConvPoolLayer(rng, input = layer_1.output, image_shape = (batch_size, nkerns[1],6,6),
filter_shape = (nkerns[2],nkerns[1],3,3), poolsize=(2,2) )
# make the input to hidden layer 2 dimensional
layer_3_input = layer_2.output.flatten(2)
layer_3 = HiddenLayer(rng,input = layer_3_input, n_in = nkerns[2]*2*2, n_out = 200, activation = T.tanh)
layer_4 = LogReg(input = layer_3.output, n_in=200, n_out = 10)
teacher_p_y_given_x = theano.shared(numpy.asarray(pickle.load(open('prob_best_model.pkl','rb')),dtype =theano.config.floatX), borrow=True)
#cost = layer_4.neg_log_likelihood(y) + T.mean((teacher_W - layer_4.W)**2)/(2.*(1+epoch*2)) + T.mean((teacher_b-layer_4.b)**2)/(2.*(1+epoch*2))
# import pdb
# pdb.set_trace()
p_y_given_x = T.matrix('p_y_given_x')
e = theano.shared(value = 0, name = 'e', borrow = True)
#cost = layer_4.neg_log_likelihood(y) + 1.0/(e)*T.mean((layer_4.p_y_given_x - p_y_given_x)**2)
cost = layer_4.neg_log_likelihood(y) + 2.0/(e)*T.mean(-T.log(layer_4.p_y_given_x)*p_y_given_x - layer_4.p_y_given_x*T.log(p_y_given_x))
test_model = theano.function([index],layer_4.errors(y),
givens={
x: test_set_x[index*batch_size:(index+1)*batch_size],
y: test_set_y[index*batch_size:(index+1)*batch_size]
})
validate_model = theano.function([index],layer_4.errors(y),
givens={
x: valid_set_x[index*batch_size:(index+1)*batch_size],
y: valid_set_y[index*batch_size:(index+1)*batch_size]
})
# list of parameters
params = layer_4.params + layer_3.params + layer_2.params + layer_1.params + layer_0.params
grads = T.grad(cost,params)
updates = [ (param_i, param_i-learning_rate*grad_i) for param_i, grad_i in zip(params,grads) ]
train_model = theano.function([index,train_epoch],cost, updates=updates,
givens={
x: train_set_x[index*batch_size:(index+1)*batch_size],
y: train_set_y[index*batch_size:(index+1)*batch_size],
p_y_given_x: teacher_p_y_given_x[index],
e: train_epoch
})
# -----------------------------------------Starting Training ------------------------------
print ('..... Training ' )
# for early stopping
patience = 10000
patience_increase = 2
improvement_threshold = 0.95
validation_frequency = min(n_train_batches, patience//2)
best_validation_loss = numpy.inf # initialising loss to be inifinite
best_itr = 0
test_score = 0
start_time = timeit.default_timer()
#epo = theano.shared('epo')
epoch = 0
done_looping = False
while (epoch < n_epochs) and (not done_looping):
epoch = epoch+1
for minibatch_index in range(n_train_batches):
iter = (epoch - 1)*n_train_batches + minibatch_index
if iter%100 ==0:
print ('training @ iter = ', iter)
cost_ij = train_model(minibatch_index,epoch)
if(iter +1)%validation_frequency ==0:
# compute loss on validation set
validation_losses = [validate_model(i) for i in range(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
# import pdb
# pdb.set_trace()
print ('epoch %i, minibatch %i/%i, validation error %f %%' %(epoch,minibatch_index+1,n_train_batches,this_validation_loss*100. ))
# check with best validation score till now
if this_validation_loss<best_validation_loss:
# improve
if this_validation_loss < best_validation_loss * improvement_threshold:
patience = max(patience, iter*patience_increase)
best_validation_loss = this_validation_loss
best_itr = iter
test_losses = [test_model(i) for i in range(n_test_batches)]
test_score = numpy.mean(test_losses)
print ('epoch %i, minibatch %i/%i, testing error %f %%' %(epoch, minibatch_index+1,n_train_batches,test_score*100.))
with open('best_model_3layer.pkl', 'wb') as f:
pickle.dump(params, f)
with open('Results_student_3.txt','wb') as f2:
f2.write(str(test_score*100) + '\n')
#if patience <= iter:
# done_looping = True
# break
end_time = timeit.default_timer()
print ('Optimization complete')
print ('Best validation score of %f %% obtained at iteration %i,'
'with test performance %f %%' % (best_validation_loss*100., best_itr, test_score*100 ))
print('The code ran for %.2fm' %((end_time - start_time)/60.))
if __name__ == '__main__':
evaluate_lenet5()