-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathdeep_net.py
332 lines (268 loc) · 13 KB
/
deep_net.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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
from __future__ import print_function
__docformat__ = 'restructedtext en'
import six.moves.cPickle as pickle
import gzip
import os
import sys
import timeit
import time
import numpy as np
import chromosome
import gene
import random
import matenc
import tensorflow as tf
def func(last, current):
return [last[0] + 1, current]
def sigmoid(arr):
return 1 / (1 + np.exp(-arr))
def relu(arr):
return np.where(arr>0,arr,0)
def dense_to_sparse(mat):
idx = tf.where(tf.not_equal(mat, 0))
sparse = tf.SparseTensor(idx, tf.gather_nd(mat, idx), mat.get_shape())
return sparse
def find_density(mat):
idx = tf.where(tf.not_equal(mat, 0))
#val = tf.gather_nd(mat, idx)
tup = tf.shape(mat)
return tf.divide( tf.shape(idx)[0],tf.multiply(tup[0],tup[1]))
def opt_compwise_multiply(mat1,mat2):
tempnode = dense_to_sparse( mat1 ).__mul__( mat2 )
return tf.scatter_nd(tempnode.indices, tempnode.values, tempnode.dense_shape)
class DeepNet(object):
def __init__(self, inputh, n_in, n_out, mat_enc, middle_activation = tf.nn.relu, final_activation = tf.nn.sigmoid):
self.n_in = n_in
self.n_out = n_out
#self.chromo = chromo
self.mat_enc = mat_enc
self.input = inputh
self.con_mat_var_map= {}
self.wei_mat_var_map = {}
for key in self.mat_enc.CMatrix.keys():
self.con_mat_var_map[key] = tf.Variable(initial_value= self.mat_enc.CMatrix[key].astype('float32'), name = 'con_mat'+key, dtype = tf.float32)
self.wei_mat_var_map[key] = tf.Variable( initial_value = self.mat_enc.WMatrix[key], name = 'con_mat'+key, dtype = tf.float32)
to_effec_mat_node_map = {}
for key in self.con_mat_var_map.keys():
to_effec_mat_node_map[key] = opt_compwise_multiply(self.con_mat_var_map[ key ],self.wei_mat_var_map[ key ])
density_map = {}
for key in to_effec_mat_node_map.keys():
density_map[key] = find_density(to_effec_mat_node_map[key])
self.bias_wei_arr = np.array( [ item.weight for item in self.mat_enc.Bias_conn_arr] )
self.bias_var = tf.Variable( initial_value = self.bias_wei_arr, name = "bias", dtype = tf.float32)
input_till_H2 = None
if 'IH1' in to_effec_mat_node_map.keys():
input_till_H1 = middle_activation(tf.sparse_matmul( self.input, to_effec_mat_node_map['IH1'], b_is_sparse = True))
if 'IH2' in to_effec_mat_node_map.keys():
input_till_H2 = tf.sparse_matmul(self.input, to_effec_mat_node_map['IH2'], b_is_sparse = True)
if 'H1H2' in to_effec_mat_node_map.keys():
assert( 'IH1' in to_effec_mat_node_map.keys())
twoh = tf.sparse_matmul( input_till_H1, to_effec_mat_node_map['H1H2'], b_is_sparse = True)
if 'IH2' in to_effec_mat_node_map.keys():
input_till_H2 = tf.add(twoh, input_till_H2)
else:
input_till_H2 = twoh
if input_till_H2 is not None:
input_till_H2 = middle_activation(input_till_H2)
output = None
if 'H2O' in to_effec_mat_node_map.keys():
assert('IH2' in to_effec_mat_node_map.keys() or 'H1H2' in to_effec_mat_node_map.keys())
threeh = tf.sparse_matmul(input_till_H2, to_effec_mat_node_map['H2O'], b_is_sparse=True)
output = threeh
if 'H1O' in to_effec_mat_node_map.keys():
assert ('IH1' in to_effec_mat_node_map.keys())
fourh = tf.sparse_matmul(input_till_H1, to_effec_mat_node_map['H1O'], b_is_sparse = True)
if output is not None:
output = tf.add( output, fourh)
else:
output = fourh
if 'IO' in to_effec_mat_node_map.keys():
assert ('IO' in to_effec_mat_node_map.keys())
fifth = tf.sparse_matmul(self.input, to_effec_mat_node_map['IO'] , b_is_sparse = True)
if output is not None:
output = tf.add( output, fifth)
else:
output = fifth
output = final_activation(output)
"""input_till_H2 = middle_activation(
tf.add(
tf.sparse_matmul(self.input, to_effec_mat_node_map['IH2'], b_is_sparse = True),
tf.sparse_matmul( input_till_H1, to_effec_mat_node_map['H1H2'], b_is_sparse = True)
)
)
output = final_activation(
tf.add(
tf.add(
tf.add(
tf.sparse_matmul(input_till_H2, to_effec_mat_node_map['H2O'], b_is_sparse = True ),
tf.sparse_matmul(input_till_H1,to_effec_mat_node_map['H1O'], b_is_sparse = True)
),
tf.sparse_matmul(self.input, to_effec_mat_node_map['IO'] , b_is_sparse = True)
),
self.bias_var
)
)
"""
self.p_y_given_x = output
half = tf.constant(0.5, dtype=self.p_y_given_x.dtype)
if int(self.bias_wei_arr.shape[0]) != 1:
self.y_pred = tf.argmax(self.p_y_given_x, axis=1)
else:
half = tf.constant(0.5, dtype=self.p_y_given_x.dtype)
dadum = tf.constant(0.5, dtype=self.p_y_given_x.dtype)
q = tf.scan(lambda last, current: current[0], elems=self.p_y_given_x, initializer=dadum)
s = tf.scan(lambda y, x: tf.greater_equal(x, half), elems=q, initializer=False)
#print("herehrerhehrehrehrehrhe", s)
# print("hi",s)
self.y_pred = tf.cast(s, dtype=tf.int32)
self.params = [ self.wei_mat_var_map[key] for key in self.wei_mat_var_map.keys()] + [self.bias_var]
def negative_log_likelihood(self, y):
if int(self.bias_wei_arr.shape[0])!=1:
dum=tf.constant(0.5,dtype=tf.float32) #dum for dummy
dadum=tf.constant(-1,dtype=tf.int32)# dum-dadum-dadum mast h
q=tf.scan(fn=func,elems=y,initializer=[dadum,dadum])
z=tf.transpose(tf.stack([q[0],q[1]]))
#print("hello---------------------------")
w=tf.scan(lambda last,current: tf.log(self.p_y_given_x[current[0]][current[1]]),elems=z, initializer = dum)
#print(-tf.reduce_mean(w))
return -tf.reduce_mean(w)
else:
dum=tf.constant(0.5,dtype=tf.float64)
minusone=tf.constant(-1,dtype=tf.int32)
one=tf.constant(1,dtype=y.dtype)
r=tf.scan(lambda last,current:last+1,elems=y,initializer=minusone)
w=tf.scan(lambda last,current: tf.add(tf.multiply(tf.cast(y[current],dtype=self.p_y_given_x.dtype),tf.log(self.p_y_given_x[current][0])),tf.multiply(tf.cast(tf.add(one,-y[current]),dtype=self.p_y_given_x.dtype),tf.log(tf.add(tf.cast(one,dtype=self.p_y_given_x.dtype),-self.p_y_given_x[current][0])))),elems=r,initializer=dum)
z=-tf.reduce_mean(w)
return z
def errors(self, y):
"""Return a float representing the number of errors in the minibatch
over the total number of examples of the minibatch ; zero one
loss over the size of the minibatch
:type y: theano.tensor.TensorType
:param y: corresponds to a vector that gives for each example the
correct label
"""
# check if y has same dimension of y_pred
"""if len(y.shape) != len(self.y_pred.shape):
raise TypeError(
'y should have the same shape as self.y_pred',
('y', y.type, 'y_pred', self.y_pred.type)
)
"""
# check if y is of the correct datatype
if y.dtype:
# the T.neq operator returns a vector of 0s and 1s, where 1
# represents a mistake in prediction
r = tf.scan(lambda last, current: last + 1, elems=y, initializer=-1)
qn = tf.scan(lambda last, current: tf.not_equal(tf.cast(self.y_pred[current], dtype=tf.int32), y[current]),
elems=r, initializer=False)
q = tf.cast(qn, dtype=tf.int32)
# r=tf.scan((lambda last,current: current[1]),q)
return tf.reduce_mean(tf.cast(q, dtype=tf.float64))
else:
raise NotImplementedError()
def test2():
for_node = [(i, 'I') for i in range(1, 4)]
for_node += [(i, 'O') for i in range(4, 6)]
st='2212211'
for_node += [(i+6,'H'+st[i]) for i in range(len(st))]
node_ctr = 13
innov_num = 25
dob = 0
node_lis = [gene.Node(x, y) for x, y in for_node]
for_conn = [(1, (1, 4), 0.3, True), (2, (1, 5), 0.25, False), (3, (2, 4), 0.25, False), (4, (2, 5), 0.5, False),
(5, (3, 4), 0.7, False), (6, (3, 5), 0.5, True), (7, (1, 6), 0.2, True), (8, (6, 4), 0.1, True),
(9, (2, 7), 0.1, True), (10, (7, 4), 0.15, True), (11, (1, 8), 0.5, True), (12, (8, 6), 0.7, True),
(13, (1, 9), 0.3, False), (14, (9, 5), 1.0, True), (15, (3, 10), 0.33, True), (16, (10, 5), 0.77, True),
(17, (1, 11), 0.25, True), (18, (11, 9), 0.15, True), (19, (2, 12), 0.6, True), (20, (12, 7), 0.4, True),
(21, (3, 12), 0.8, True), (22, (2, 9), 0.9, True), (23, (12, 4), 0.75, True), (24, (11, 5), 0.25, True),
]
conn_lis = [gene.Conn(x, (node_lis[tup[0] - 1], node_lis[tup[1] - 1]), w, status) for x, tup, w, status in for_conn]
for_bias = [(4, 0.2), (5, 0.1)]
bias_conn_lis = [gene.BiasConn(node_lis[x - 1], y) for x, y in for_bias]
newchromo = chromosome.Chromosome(dob, node_lis, conn_lis, bias_conn_lis)
newchromo.set_node_ctr(node_ctr)
# newchromo.pp()
def calc_output_directly(inputarr):
lis = []
for arr in inputarr:
x1 = arr[0]
x2 = arr[1]
x3 = arr[2]
output1 = sigmoid(
0.3 * x1 +
0.1 * relu(
0.7 * relu(0.5 * x1) +
0.2 * x1
) +
0.15 * relu(
0.1 * x2 +
0.4 * relu(
0.6 * x2 +
0.8 * x3
)
) +
0.75 * relu(
0.6 * x2 +
0.8 * x3
) -
0.2
)
#output2 = sigmoid(arr[0] * 0.25 + arr[1] * 0.5 + relu(arr[2] * 0.3) * 0.6 - 0.1)
output2 = sigmoid(
0.5 * x3 +
1 * relu(
0.15 * relu(0.25 * x1) +
0.9 * x2
) +
0.25 * relu(
0.25 * x1
) +
0.77 * relu(
0.33 * x3
) -
0.1
)
lis.append([output1, output2])
return np.array(lis)
inputarr = np.array([[0.0, 2, 1], [0.8, 1, 2]])
indim = 3
outdim = 2
#np.random
rng = random
num_data = 2
# inputarr = np.random.random((num_data, indim))
#neter = Neterr(indim, outdim, inputarr, 10, np.random)
ka = random.randint(0,2,(num_data,))
"""
targetarr = np.zeros((num_data,outdim)).astype(dtype = 'float32')
for i in range(num_data):
targetarr[i,ka[i]] = 1
print("target is ", targetarr)
"""
targetarr = ka.astype('int32')
print(targetarr.dtype)
inputarr = inputarr.astype('float32')
print("input type", inputarr.dtype)
print(targetarr)
x = tf.placeholder( shape = [None, indim], dtype = tf.float32)
y = tf.placeholder( shape = [None,], dtype = tf.int32)
newmat_enc = newchromo.convert_to_MatEnc(indim,outdim)
newnet = DeepNet(x, indim, outdim, newmat_enc)
cost = newnet.negative_log_likelihood(y)
learning_rate = 0.05
optmzr = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost, var_list=newnet.params)
#cost = newnet.errors(y)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print("initially")
print(sess.run(newnet.wei_mat_var_map['IH2']))
print(sess.run(newnet.con_mat_var_map['IH2']))
print(sess.run([optmzr,newnet.y_pred,cost], feed_dict = { x : inputarr, y : targetarr}))
print(sess.run(newnet.wei_mat_var_map['IH2']))
print(sess.run([optmzr, newnet.bias_var, cost], feed_dict={x: inputarr, y: targetarr}))
print(sess.run(newnet.con_mat_var_map['IH2']))
print(sess.run([optmzr, newnet.bias_var, newnet.errors(y)], feed_dict={x: inputarr, y: targetarr}))
#newchromo.modify_thru_backprop()
if __name__ == '__main__':
test2()