-
Notifications
You must be signed in to change notification settings - Fork 34
/
recGenI64.py
162 lines (120 loc) · 7.16 KB
/
recGenI64.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
''' Version 1.000
Code provided by Daniel Jiwoong Im and Chris Dongjoo Kim
Permission is granted for anyone to copy, use, modify, or distribute this
program and accompanying programs and documents for any purpose, provided
this copyright notice is retained and prominently displayed, along with
a note saying that the original programs are available from our
web page.
The programs and documents are distributed without any warranty, express or
implied. As the programs were written for research purposes only, they have
not been tested to the degree that would be advisable in any important
application. All use of these programs is entirely at the user's own risk.'''
'''Demo of Generating images with recurrent adversarial networks.
For more information, see: http://arxiv.org/abs/1602.05110
'''
import os
import sys
import theano
import theano.tensor as T
import numpy as np
import scipy as sp
from batch_norm_conv_layer import *
from batch_norm_layer import *
from conv_layer import *
from deconv import *
from utils import *
from collections import OrderedDict
import theano.sandbox.rng_mrg as RNG_MRG
rng = np.random.RandomState()
MRG = RNG_MRG.MRG_RandomStreams(rng.randint(2 ** 30))
class RecGenI64(object):
def __init__(self, model_params, nkerns=[1,8,4,2,1], ckern=172, filter_sizes=[4,5,5,5,5]):
"""Initialize the architecture of the model"""
[self.batch_sz, num_dims, self.num_hids, numpy_rng,\
self.dim_sample, nkerns, ckern, self.num_channel, self.num_steps] = model_params
self.nkerns = np.asarray(nkerns) * ckern # of constant gen filters in first conv layer
self.nkerns[-1] = 3
self.D = int(np.sqrt(num_dims / self.nkerns[-1]))
self.numpy_rng = numpy_rng
self.filter_sizes=filter_sizes
self.L0a = Batch_Norm_layer(self.dim_sample, self.dim_sample, 'W_h_z', numpy_rng)
self.L1 = Batch_Norm_layer(self.dim_sample*2, self.num_hids[0], 'W_h_hf', numpy_rng)
self.L0b = Batch_Norm_layer(self.num_hids[0], self.dim_sample, 'W_z_f', numpy_rng)
self.L2 = Deconv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[1], bnkern=self.nkerns[0] , bfilter_sz=filter_sizes[0], tfilter_sz=filter_sizes[1])
self.L3 = Deconv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[2], bnkern=self.nkerns[1] , bfilter_sz=filter_sizes[1], tfilter_sz=filter_sizes[2])
self.L4 = Deconv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[3], bnkern=self.nkerns[2] , bfilter_sz=filter_sizes[2], tfilter_sz=filter_sizes[3])
self.L5 = Deconv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[4], bnkern=self.nkerns[3] , bfilter_sz=filter_sizes[3], tfilter_sz=filter_sizes[4])
self.L6 = Conv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[1], bnkern=self.nkerns[0] , bfilter_sz=filter_sizes[0], tfilter_sz=filter_sizes[1])
self.L7 = BN_Conv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[2], bnkern=self.nkerns[1] , bfilter_sz=filter_sizes[1], tfilter_sz=filter_sizes[2])
self.L8 = BN_Conv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[3], bnkern=self.nkerns[2] , bfilter_sz=filter_sizes[2], tfilter_sz=filter_sizes[3])
self.L9 = BN_Conv_layer(self.batch_sz, numpy_rng, tnkern=self.nkerns[4], bnkern=self.nkerns[3] , bfilter_sz=filter_sizes[3], tfilter_sz=filter_sizes[4])
self.params = self.L0a.params + self.L0b.params + self.L1.params + \
self.L2.params + self.L3.params + self.L4.params \
+ self.L5.params + self.L6.params + self.L7.params \
+ self.L8.params +self.L9.params
def init_params(self, numpy_rng):
"""initialize the weights and biases to be used throughout the model"""
self.W_z_f = initialize_weight(self.num_hids[0], self.dim_sample, 'W_z_f', numpy_rng, 'uniform')
self.z_f_bias = theano.shared(np.zeros((self.dim_sample,), dtype=theano.config.floatX), name='z_f_bias')
self.W_h_z = initialize_weight(self.dim_sample, self.dim_sample, 'W_h_z', numpy_rng, 'uniform')
self.h_z_bias = theano.shared(np.zeros((self.dim_sample,), dtype=theano.config.floatX), name='h_z_bias')
self.W_h_hf = initialize_weight(self.dim_sample*2, self.num_hids[0], 'W_h_hf', numpy_rng, 'uniform')
self.h_hf_bias = theano.shared(np.zeros((self.num_hids[0],), dtype=theano.config.floatX), name='h_hf_bias')
self.param0 = [self.W_h_z, self.h_z_bias, self.W_z_f, self.z_f_bias, self.W_h_hf, self.h_hf_bias]
def forward(self, Z, H_Ct, testF=False, atype='relu'):
"""
The forward propagation to generate the C_i image.
"""
H0 = self.L0a.propagate(Z, testF=testF, atype='tanh')
R0 = T.concatenate([H0.T, H_Ct.T]).T
H1 = self.L1.propagate( R0, testF=testF, atype=atype)
H1 = H1.reshape((R0.shape[0], self.nkerns[0], self.filter_sizes[0], self.filter_sizes[0]))
H2 = self.L2.deconv(H1, atype=atype)
H3 = self.L3.deconv(H2, atype=atype)
H4 = self.L4.deconv(H3, atype=atype)
H5 = self.L5.deconv(H4, atype='linear')
return H5
def backward(self,Ct, atype='relu', testF=False):
"""
backprop or "inverse" stage to update h_i
"""
H0 = self.L9.conv(Ct, atype=atype)
H1 = self.L8.conv(H0, atype=atype)
H2 = self.L7.conv(H1, atype=atype)
H3 = self.L6.conv(H2, atype=atype)
H3 = H3.flatten(2)
return self.L0b.propagate(H3, testF=testF, atype='tanh')
def get_samples(self, num_sam, scanF=True):
"""
Retrieves the samples for the current time step.
uncomment parts when time step changes.
"""
print 'Get_sample func: Number of steps iterate over ::: %d' % self.num_steps
H_Ct = T.alloc(0., num_sam, self.dim_sample)
Zs = MRG.normal(size=(self.num_steps, num_sam, self.dim_sample), avg=0., std=1.)
Canvases = self.apply_recurrence(self.num_steps, Zs, H_Ct)
C = T.sum(T.stacklists(Canvases),axis=0)
return activation_fn_th(C, atype='sigmoid'), Canvases
def apply_recurrence(self, num_steps, Zs, H_Ct, scanF=True):
"""
Applies the recurrent architecture of Model described in GRAN paper, to recursively
input previous step's encoded Canvas(H_Ct) with random noise(Z), and receive the
encoded H_Ct for the current time step
It returns the final output Canvas, C.
"""
def recurrence(i, H_Ct, Zs):
Z = Zs[0]
Ct = self.forward(Z, H_Ct)
H_Ct = self.backward(Ct)
return H_Ct, Ct
if scanF:
[H_Ct, C], updates = theano.scan(fn=recurrence, outputs_info=[H_Ct, None], \
sequences=[T.arange(num_steps, dtype='int16')], non_sequences=[Zs])
else:
for i in xrange(num_steps):
H_Ct, C = recurrence(i, H_Ct, C, Zs)
return C
def weight_decay_l2(self):
"""l2 weight decay used in the optimize_gan for computing the cost of the discriminator"""
return 0.5*((self.L2.W **2).sum() \
+ (self.L3.W**2).sum() + (self.L4.W**2).sum() + (self.L5.W**2).sum())