-
Notifications
You must be signed in to change notification settings - Fork 1
/
cell.py
209 lines (176 loc) · 7.47 KB
/
cell.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
import tensorflow as tf
from spatial_transformer import ElasticTransformer
from model import *
class ConvLSTMCell(tf.nn.rnn_cell.RNNCell):
"""A LSTM cell with convolutions instead of multiplications.
Reference:
Xingjian, S. H. I., et al. "Convolutional LSTM network: A machine learning approach for precipitation nowcasting." Advances in Neural Information Processing Systems. 2015.
"""
def __init__(self, shape, filters, kernel, forget_bias=1.0, activation=tf.tanh, normalize=True, peephole=True, data_format='channels_last', reuse=None):
super(ConvLSTMCell, self).__init__(_reuse=reuse)
self._kernel = kernel
self._filters = filters
self._forget_bias = forget_bias
self._activation = activation
self._normalize = normalize
self._peephole = peephole
if data_format == 'channels_last':
self._size = tf.TensorShape(shape + [self._filters])
self._feature_axis = self._size.ndims
self._data_format = None
elif data_format == 'channels_first':
self._size = tf.TensorShape([self._filters] + shape)
self._feature_axis = 0
self._data_format = 'NC'
else:
raise ValueError('Unknown data_format')
@property
def state_size(self):
return tf.nn.rnn_cell.LSTMStateTuple(self._size, self._size)
@property
def output_size(self):
return self._size
def call(self, x, state):
c, h = state
x = tf.concat([x, h], axis=self._feature_axis)
n = x.shape[-1].value
m = 4 * self._filters if self._filters > 1 else 4
W = tf.get_variable('kernel', self._kernel + [n, m])
y = tf.nn.convolution(x, W, 'SAME', data_format=self._data_format)
if not self._normalize:
y += tf.get_variable('bias', [m], initializer=tf.zeros_initializer())
j, i, f, o = tf.split(y, 4, axis=self._feature_axis)
if self._peephole:
i += tf.get_variable('W_ci', c.shape[1:]) * c
f += tf.get_variable('W_cf', c.shape[1:]) * c
if self._normalize:
j = tf.contrib.layers.layer_norm(j)
i = tf.contrib.layers.layer_norm(i)
f = tf.contrib.layers.layer_norm(f)
f = tf.sigmoid(f + self._forget_bias)
i = tf.sigmoid(i)
c = c * f + i * self._activation(j)
if self._peephole:
o += tf.get_variable('W_co', c.shape[1:]) * c
if self._normalize:
o = tf.contrib.layers.layer_norm(o)
c = tf.contrib.layers.layer_norm(c)
o = tf.sigmoid(o)
h = o * self._activation(c)
state = tf.nn.rnn_cell.LSTMStateTuple(c, h)
return h, state
class ConvGRUCell(tf.nn.rnn_cell.RNNCell):
"""A GRU cell with convolutions instead of multiplications."""
def __init__(self, shape, filters, kernel, activation=tf.tanh, normalize=True, data_format='channels_last', reuse=None):
super(ConvGRUCell, self).__init__(_reuse=reuse)
self._filters = filters
self._kernel = kernel
self._activation = activation
self._normalize = normalize
if data_format == 'channels_last':
self._size = tf.TensorShape(shape + [self._filters])
self._feature_axis = self._size.ndims
self._data_format = None
elif data_format == 'channels_first':
self._size = tf.TensorShape([self._filters] + shape)
self._feature_axis = 0
self._data_format = 'NC'
else:
raise ValueError('Unknown data_format')
@property
def state_size(self):
return self._size
@property
def output_size(self):
return self._size
def call(self, x, h):
channels = x.shape[self._feature_axis].value
with tf.variable_scope('gates'):
inputs = tf.concat([x, h], axis=self._feature_axis)
n = channels + self._filters
m = 2 * self._filters if self._filters > 1 else 2
W = tf.get_variable('kernel', self._kernel + [n, m])
y = tf.nn.convolution(inputs, W, 'SAME', data_format=self._data_format)
if self._normalize:
r, u = tf.split(y, 2, axis=self._feature_axis)
r = tf.contrib.layers.layer_norm(r)
u = tf.contrib.layers.layer_norm(u)
else:
y += tf.get_variable('bias', [m], initializer=tf.ones_initializer())
r, u = tf.split(y, 2, axis=self._feature_axis)
r, u = tf.sigmoid(r), tf.sigmoid(u)
with tf.variable_scope('candidate'):
inputs = tf.concat([x, r * h], axis=self._feature_axis)
n = channels + self._filters
m = self._filters
W = tf.get_variable('kernel', self._kernel + [n, m])
y = tf.nn.convolution(inputs, W, 'SAME', data_format=self._data_format)
if self._normalize:
y = tf.contrib.layers.layer_norm(y)
else:
y += tf.get_variable('bias', [m], initializer=tf.zeros_initializer())
h = u * h + (1 - u) * self._activation(y)
return h, h
# UNet_down_template = tf.make_template('UNet_down', UNet_down)
# class ConvGRUCell(tf.nn.rnn_cell.RNNCell):
# """A GRU cell with convolutions instead of multiplications."""
# def __init__(self, shape, filters, kernel, config, is_train, activation=tf.tanh, normalize=True, data_format='channels_last', reuse=None):
# super(ConvGRUCell, self).__init__(_reuse=reuse)
# self._filters = filters
# self._kernel = kernel
# self._activation = activation
# self._normalize = normalize
# self.stl = ElasticTransformer(config.out_size, config.grid_size)
# self.config = config
# self.is_train = is_train
# if data_format == 'channels_last':
# #self._size = tf.TensorShape(shape + [self._filters])
# self._size = tf.TensorShape(shape + [3])
# self._feature_axis = self._size.ndims
# self._data_format = None
# elif data_format == 'channels_first':
# self._size = tf.TensorShape([self._filters] + shape)
# #self._size = tf.TensorShape([3] + shape)
# self._feature_axis = 0
# self._data_format = 'NC'
# else:
# raise ValueError('Unknown data_format')
# @property
# def state_size(self):
# return self._size
# @property
# def output_size(self):
# return self._size
# def call(self, x, h):
# x_origin = x
# with tf.variable_scope('gates'):
# x = UNet_down_template(x, is_train = True) # 22 X 22 X 256
# h = UNet_down_template(h, is_train = True) # 22 X 22 X 256
# channels = x.shape[self._feature_axis].value
# inputs = tf.concat([x, h], axis=self._feature_axis)
# n = channels + self._filters
# m = 2 * self._filters if self._filters > 1 else 2
# W = tf.get_variable('kernel', self._kernel + [n, m])
# y = tf.nn.convolution(inputs, W, 'SAME', data_format=self._data_format)
# if self._normalize:
# r, u = tf.split(y, 2, axis=self._feature_axis)
# r = tf.contrib.layers.layer_norm(r)
# u = tf.contrib.layers.layer_norm(u)
# else:
# y += tf.get_variable('bias', [m], initializer=tf.ones_initializer())
# r, u = tf.split(y, 2, axis=self._feature_axis)
# r, u = tf.sigmoid(r), tf.sigmoid(u)
# with tf.variable_scope('candidate'):
# inputs = tf.concat([x, r * h], axis=self._feature_axis)
# n = channels + self._filters
# m = self._filters
# W = tf.get_variable('kernel', self._kernel + [n, m])
# y = tf.nn.convolution(inputs, W, 'SAME', data_format=self._data_format)
# if self._normalize:
# y = tf.contrib.layers.layer_norm(y)
# else:
# y += tf.get_variable('bias', [m], initializer=tf.zeros_initializer())
# h = u * h + (1 - u) * self._activation(y)
# theta = Localizer(h, self.config.grid_size, is_train = self.is_train, scope = 'localizer')
# h = self.stl.transform(x_origin, theta)
# return h, h