This repository has been archived by the owner on Jun 16, 2019. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathkeras_model.py
190 lines (155 loc) · 11.5 KB
/
keras_model.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
keras_dir = '/home/bill/Libraries/keras/'
import sys
sys.path.append(keras_dir)
from keras.models import *
from keras.layers.core import *
from keras.layers.convolutional import *
import numpy as np
class ConvLSTMLadderNet(Graph):
def __init__(self, config, build=True):
super(ConvLSTMLadderNet, self).__init__()
self.config = config
self.initialize()
if build:
self.build()
def build(self):
loss = {}
loss_weights = {}
for i,t in enumerate(self.config.t_predict):
loss['output_t%d' % t] = self.config.loss
if hasattr(self.config, 'loss_weights'):
loss_weights['output_t%d' % t] = self.config.loss_weights[i]
else:
loss_weights['output_t%d' % t] = 1.0
self.compile(loss=loss, optimizer=self.config.optimizer, loss_weights=loss_weights)
def format_data(self, X, Y=None):
data = {}
n = X.shape[0]
for t in range(self.config.nt_in):
data['input_t%d' % t] = X[:,t]
for l in range(self.config.n_modules):
data['H_l%d_t-1' % l] = np.zeros((n, self.config.stack_sizes[l], self.config.input_shape[1] // 2**(l+1), self.config.input_shape[2] // 2**(l+1))).astype(np.float32)
data['C_l%d_t-1' % l] = np.copy(data['H_l%d_t-1' % l])
if Y is not None:
for i in range(len(self.config.t_predict)):
if len(Y.shape) == 5:
data['output_t%d' % t] = Y[:,i]
else:
data['output_t%d' % t] = Y
return data
def format_predictions(self, data):
for i,t in enumerate(self.config.t_predict):
Xt = data['output_t%d' % t]
if i==0:
X = np.zeros( (Xt.shape[0], len(self.config.t_predict)) + Xt.shape[1:]).astype(np.float32)
X[:,i] = Xt
return X
def initialize(self):
# initialize hidden states
for l in range(self.config.n_modules):
self.add_input(name='H_l%d_t-1' % l, input_shape=(self.config.stack_sizes[l], self.config.input_shape[0] // 2**(l+1), self.config.input_shape[1] // 2**(l+1)))
self.add_input(name='C_l%d_t-1' % l, input_shape=(self.config.stack_sizes[l], self.config.input_shape[0] // 2**(l+1), self.config.input_shape[1] // 2**(l+1)))
for t in range(self.config.nt_in):
self.add_input(name='input_t%d' % t, input_shape=(1, self.config.input_shape[0], self.config.input_shape[1]))
# add first conv layer
if t == 0:
trainable = True
shared_layer = None
else:
trainable = False
shared_layer = self.nodes['conv0pre_l-1_t0']
layer = Convolution2D(self.config.stack_sizes[0], 5, 5, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layer)
self.add_node(layer, name='conv0pre_l-1_t%d' % t, input='input_t%d' % t)
self.add_node(AveragePooling2D(), name='conv0_l-1_t%d' % t, input='conv0pre_l-1_t%d' % t)
for l in range(self.config.n_modules):
layer_names = ['conv%d_l%d_t%d' % (i, l, t) for i in range(4)]
if t == 0:
trainable = True
shared_layers = [None for _ in range(4)]
else:
trainable = False
shared_layers = [self.nodes['conv%d_l%d_t0' % (i, l)] for i in range(4)]
if l==0:
module_input = 'conv0_l-1_t%d' % t
module_input_upchannel = 'conv0_l-1_t%d' % t
else:
module_input = 'H_l%d_t%d' % (l-1, t)
module_input_upchannel = 'Hupchannel_l%d_t%d' % (l-1, t)
if t == 0:
shared_l = None
else:
shared_l = self.nodes['Hupchannel_l%d_t0' % (l-1)]
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_l)
self.add_node(layer, name='Hupchannel_l%d_t%d' % (l-1, t), input=module_input)
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[0])
self.add_node(layer, name=layer_names[0], input=module_input)
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[1])
self.add_node(layer, name=layer_names[1], input=layer_names[0])
self.add_node(Activation('linear'), name='res0_l%d_t%d' % (l, t), inputs=[module_input_upchannel, layer_names[1]], merge_mode='sum')
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[2])
self.add_node(layer, name=layer_names[2], input='res0_l%d_t%d' % (l, t))
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[3])
self.add_node(layer, name=layer_names[3], input=layer_names[2])
if l > 0:
self.add_node(AveragePooling2D(), name='res1_l%d_t%d' % (l, t), inputs=['res0_l%d_t%d' % (l, t), layer_names[3]], merge_mode='sum')
else:
self.add_node(Activation('linear'), name='res1_l%d_t%d' % (l, t), inputs=['res0_l%d_t%d' % (l, t), layer_names[3]], merge_mode='sum')
if t==0:
shared_layers = [None for _ in range(4)]
else:
shared_layers = [self.nodes['I_l%d_t0' % l], self.nodes['F_l%d_t0' % l], self.nodes['O_l%d_t0' % l], self.nodes['C1_l%d_t0' % l] ]
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='hard_sigmoid', trainable=trainable, shared_layer=shared_layers[0])
self.add_node(layer, name='I_l%d_t%d' % (l, t), inputs=['res1_l%d_t%d' % (l, t), 'H_l%d_t%d' % (l, t-1)], merge_mode='concat', concat_axis=-3)
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='hard_sigmoid', trainable=trainable, shared_layer=shared_layers[1])
self.add_node(layer, name='F_l%d_t%d' % (l, t), inputs=['res1_l%d_t%d' % (l, t), 'H_l%d_t%d' % (l, t-1)], merge_mode='concat', concat_axis=-3)
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='hard_sigmoid', trainable=trainable, shared_layer=shared_layers[2])
self.add_node(layer, name='O_l%d_t%d' % (l, t), inputs=['res1_l%d_t%d' % (l, t), 'H_l%d_t%d' % (l, t-1)], merge_mode='concat', concat_axis=-3)
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='tanh', trainable=trainable, shared_layer=shared_layers[3])
self.add_node(layer, name='C1_l%d_t%d' % (l, t), inputs=['res1_l%d_t%d' % (l, t), 'H_l%d_t%d' % (l, t-1)], merge_mode='concat', concat_axis=-3)
self.add_node(Activation('linear'), name='C2_l%d_t%d' % (l, t), inputs=['I_l%d_t%d' % (l, t), 'C1_l%d_t%d' % (l, t)], merge_mode='mul')
self.add_node(Activation('linear'), name='C3_l%d_t%d' % (l, t), inputs=['F_l%d_t%d' % (l, t), 'C_l%d_t%d' % (l, t-1)], merge_mode='mul')
self.add_node(Activation('linear'), name='C_l%d_t%d' % (l, t), inputs=['C3_l%d_t%d' % (l, t), 'C2_l%d_t%d' % (l, t)], merge_mode='sum')
self.add_node(Activation('tanh'), name='Ct_l%d_t%d' % (l, t), input='C_l%d_t%d' % (l, t))
self.add_node(Activation('linear'), name='H_l%d_t%d' % (l, t), inputs=['O_l%d_t%d' % (l, t), 'Ct_l%d_t%d' % (l, t)], merge_mode='mul')
if t in self.config.t_predict:
for l in range(self.config.n_modules-1, -1, -1):
layer_names = ['deconv%d_l%d_t%d' % (i, l, t) for i in range(4)]
if t == self.config.t_predict[0]:
trainable = True
shared_layers = [None for _ in range(4)]
else:
trainable = False
shared_layers = [self.nodes['deconv%d_l%d_t%d' % (i, l, self.config.t_predict[0])] for i in range(4)]
if l==self.config.n_modules-1:
module_input = 'Hup_l%d_t%d' % (l, t)
else:
module_input = 'comb_l%d_t%d' % (l, t)
self.add_node(UpSampling2D(), name='Hup_l%d_t%d' % (l, t), input='H_l%d_t%d' % (l, t))
if l<self.config.n_modules-1:
self.add_node(UpSampling2D(), name='deres1up_l%d_t%d' % (l+1, t), input='deres1_l%d_t%d' % (l+1, t))
if l<self.config.n_modules-1:
#self.add_node(Activation('linear'), name='prod_l%d_t%d' % (l, t), inputs=['Hup_l%d_t%d' % (l, t), 'deres1_l%d_t%d' % (l+1, t)], merge_mode='mul')
if t == self.config.t_predict[0]:
shared_l = None
tr = True
else:
shared_l = self.nodes['comb_l%d_t%d' % (l, self.config.t_predict[0])]
tr = False
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=tr, shared_layer=shared_l)
self.add_node(layer, name='comb_l%d_t%d' % (l, t), inputs=['Hup_l%d_t%d' % (l, t), 'deres1up_l%d_t%d' % (l+1, t)], merge_mode='concat', concat_axis=-3)
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[0])
self.add_node(layer, name=layer_names[0], input=module_input)
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[1])
self.add_node(layer, name=layer_names[1], input=layer_names[0])
self.add_node(Activation('linear'), name='deres0_l%d_t%d' % (l, t), inputs=[module_input, layer_names[1]], merge_mode='sum')
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[2])
self.add_node(layer, name=layer_names[2], input='deres0_l%d_t%d' % (l, t))
layer = Convolution2D(self.config.stack_sizes[l], 3, 3, border_mode='same', activation='relu', trainable=trainable, shared_layer=shared_layers[3])
self.add_node(layer, name=layer_names[3], input=layer_names[2])
self.add_node(Activation('linear'), name='deres1_l%d_t%d' % (l, t), inputs=['deres0_l%d_t%d' % (l, t), layer_names[3]], merge_mode='sum')
if l==0:
if self.config.predict_var == 'membrane':
self.add_node(Convolution2D(1, 1, 1), name='output_conv_t%d' % t, input='deres1_l%d_t%d' % (l, t))
self.add_node(Activation('sigmoid'), name='output_t%d' % t, input='output_conv_t%d' % t, create_output=True)
else:
self.add_node(Convolution2D(1, 1, 1), name='output_t%d' % t, input='deres1_l%d_t%d' % (l, t), create_output=True)