-
Notifications
You must be signed in to change notification settings - Fork 0
/
quassi_rnn.py
160 lines (135 loc) · 5.94 KB
/
quassi_rnn.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
import tensorflow as tf
import tensorflow.examples.tutorials.mnist as mnist
import numpy as np
from sklearn.datasets import load_digits
def reshape(x):
s = tf.reshape(x, [-1, 8, 8, 1])
d = tf.zeros([x.get_shape().as_list()[0], 8, 7, 1], dtype=tf.float32)
return tf.concat(2,[d, s])
def create_weights(T, n, m):
weights = {
"W_z" : tf.Variable(tf.random_normal([n, T, 1, m])),
"W_f" : tf.Variable(tf.random_normal([n, T, 1, m])),
"W_o" : tf.Variable(tf.random_normal([n, T, 1, m])),
"W" : tf.Variable(tf.random_normal([T*m, 10])),
"h" : tf.Variable(tf.zeros([batch_size, T, m, 1], dtype = tf.float32)),
"c" : tf.Variable(tf.zeros([1, m, 1], dtype = tf.float32))
}
return weights
def update(X, weights, b, idx):
# Convolution operation
conv_z = tf.nn.conv2d(X, weights[idx]["W_z"], strides = [1, 1, 1, 1], padding = "VALID")
conv_z_reshaped = tf.nn.tanh(tf.transpose(conv_z, perm = [0,2,3,1]))
# Convolution witj forget gate params
conv_f = tf.nn.conv2d(X, weights[idx]["W_f"], strides = [1, 1, 1, 1], padding = "VALID")
conv_f_reshaped = tf.nn.sigmoid(tf.transpose(conv_f, perm = [0,2,3,1]))
# Convolution with output gate params
conv_o = tf.nn.conv2d(X, weights[idx]["W_o"], strides = [1, 1, 1, 1], padding = "VALID")
conv_o_reshaped = tf.nn.sigmoid(tf.transpose(conv_o, perm = [0,2,3,1]))
# Cell state component contributed by the input
# This doesnt need to be sequential
c_conv = (1-conv_f_reshaped)*conv_z_reshaped
# Pooling operation
c = weights[idx]["c"][0,:,:]
cell_state = []
with tf.device('/gpu:0'):
for j in range(batch_size):
ind_cell_state = []
# The following loop loops over the time dimension and generated the cell state for the
# given sample j
for i in range(c_conv.get_shape().as_list()[1]):
c = c*conv_f_reshaped[j,i,:,:] + c_conv[j, i, :, :]
ind_cell_state.append(c)
ind_cell_state = tf.convert_to_tensor(ind_cell_state)
dims_ics = ind_cell_state.get_shape().as_list()
cell_state.append(ind_cell_state)
cell_state = tf.convert_to_tensor(cell_state)
# Update to hidden state for the Minibatch
weights[idx]["h"] = conv_o_reshaped*cell_state
return c,weights[idx]["h"]
# create layers
def quasinet(X, weights, biases):
# X is of shape NxTxnx1
# N : batch size, T : width of the filter, n : no of dimensions
# layer 1 for the network
cell_state, layer_1 = update(X, weights, biases, 0)
weights[0]["h"] = layer_1
weights[0]["c"] = cell_state
# Mask the output of first layer with T-1 units
dims = weights[1]["W_z"].get_shape().as_list()
d = tf.zeros([layer_1.get_shape().as_list()[0], dims[0], dims[1]-1, 1], dtype=tf.float32)
layer_2 = tf.concat(2,[d, tf.transpose(layer_1, perm=[0,2,1,3])])
# layer 2 for the network
cell_state2,layer_2 = update(layer_2, weights, biases, 1)
weights[1]["h"] = layer_2
weights[1]["c"] = cell_state2
dim_prod = layer_2.get_shape().as_list()
# Fully connected layer for the final prediction
fc1 = tf.reshape(layer_2, [dim_prod[0], dim_prod[1]*dim_prod[2]])
fc1 = tf.nn.tanh(tf.add(tf.matmul(fc1, weights[-1]["W"]), biases[-1]))
return fc1
def next_batch(step, digits_data, digits_labels, batch_size):
"""
This function get the next batch for training
"""
X = digits_data[batch_size*(step):(step+1)*batch_size]
Y = digits_labels[batch_size*(step):(step+1)*batch_size]
return X,Y
learning_rate = 0.01
batch_size = 50
training_iters = 34*batch_size
display_step = 5
# mnist_data = mnist.input_data.read_data_sets("/tmp/data", one_hot=True)
digits = load_digits(n_class=10)
digits_data = digits["data"]
digits_labels = digits["target"]
digits_labels_oh = []
for i in range(digits_data.shape[0]):
a = [0.0]*10
a[digits_labels[i]] += 1.0
digits_labels_oh.append(a)
digits_labels = np.array(digits_labels_oh)
# X = tf.placeholder(tf.float32,[batch_size, 784])
X = tf.placeholder(tf.float32,[batch_size, 64])
X_reshaped = reshape(X)
Y = tf.placeholder(tf.float32,[batch_size, 10])
# each element is tuple T, n, m specifying the layer
# layer_params = [(28, 28, 20), (28, 20, 15)]
layer_params = [(8, 8, 18), (8, 18, 24)]
weights = []
biases = []
for i in range(len(layer_params)):
weights.append(create_weights(layer_params[i][0],layer_params[i][1], layer_params[i][2]))
biases.append(tf.Variable(tf.random_normal([10])))
pred = quasinet(X_reshaped, weights, biases)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, Y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
step = 1
# Keep training until reach max iterations
while step*batch_size < training_iters:
batch_x, batch_y = next_batch(step,digits_data, digits_labels, batch_size)
# Run optimization op (backprop)
p,q = sess.run([pred,optimizer], feed_dict={X: batch_x, Y: batch_y})
# print p[0], batch_y[0]
if step % display_step == 0:
# Calculate batch loss and accuracy
preds,loss, acc = sess.run([pred,cost, accuracy], feed_dict={X: batch_x,
Y: batch_y})
print "Iter " + str(step*batch_size) + ", Minibatch Loss= " + \
"{:.6f}".format(loss) + ", Training Accuracy= " + \
"{:.5f}".format(acc)
step += 1
print "Optimization Finished!"
# Calculate accuracy for batch_size digits
val = 0.0
for i in range(30):
s,d = next_batch(i,digits_data, digits_labels, batch_size)
f=(sess.run(accuracy, feed_dict={X: s,Y: d}))
# print f
val+=f*50
print val/(1500.0)