/
resnet.py
237 lines (202 loc) · 9.27 KB
/
resnet.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
"""Define a residual network that takes in calcium data and outputs spike
predictions."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from util.nncomponents import dense_batch_relu, \
summarize_layer, conv1d_batch_norm, scaled_xavier
from util.nnio import pad_to_batch_size
import numpy as np
import tensorflow as tf
from tensorflow.contrib import learn
from tensorflow import summary
def residual_network(features, mode, config):
seq_len = config['window_padding'] * 2 + config['window_size']
batch_size = config['batch_size']
input_layer = tf.reshape(features['X'], shape=(
batch_size, seq_len, 1), name="initial_reshape")
if config['global_features']:
global_features = features['global_features']
global_h1 = dense_batch_relu(global_features,
mode == learn.ModeKeys.TRAIN,
'global_h1',
config['N_global_intermediate_variables'])
global_h2 = dense_batch_relu(global_h1,
mode == learn.ModeKeys.TRAIN,
'global_h2',
config['N_global_latent_variables'],
use_relu=False)
global_features_processed = global_h2
global_h2 = tf.layers.dropout(global_h2,
rate=config['dropout'],
training=mode == learn.ModeKeys.TRAIN)
global_h2 = tf.nn.softmax(global_h2)
if config['global_features']:
shape = (config['conv_window_size'],
config['N_latent_variables'],
config['N_global_latent_variables'])
W1 = tf.get_variable('W1',
shape=(shape[0], 1, shape[1] * shape[2]),
initializer=tf.contrib.layers.variance_scaling_initializer(
factor=.3,
mode='FAN_AVG',
uniform=False))
# svd_weights = tf.matmul(global_h2, W1)
b1 = tf.get_variable('b1', shape=(1, 1, shape[1] * shape[2]))
h1 = tf.nn.conv1d(input_layer,
W1,
stride=1,
padding='VALID') + b1
winsize = (config['window_size'] +
2 * config['window_padding'] -
config['conv_window_size'] + 1)
h1 = tf.reshape(h1, [config['batch_size'],
winsize * config['N_latent_variables'],
config['N_global_latent_variables']])
# global
h1 = tf.einsum('aij,aj->ai', h1, global_h2)
h1 = tf.reshape(h1, [config['batch_size'],
winsize,
config['N_latent_variables']], name='h1')
else:
h1 = tf.layers.conv1d(
input_layer,
config['N_latent_variables'],
config['conv_window_size'],
strides=1,
use_bias=True,
kernel_initializer=scaled_xavier(.005),
bias_initializer=tf.constant_initializer(.01),
padding='valid',
name='h1',
activation=None)
summarize_layer('h1', h1)
h1_norm = tf.layers.batch_normalization(
h1,
axis=2,
training=(mode == learn.ModeKeys.TRAIN),
name='h1_norm')
h1_norm = tf.layers.dropout(h1_norm,
rate=config['dropout'],
training=mode == learn.ModeKeys.TRAIN)
h1_out = tf.nn.relu(h1_norm, name='h1_out')
summarize_layer('h1_out', h1_out)
sum_layer = tf.identity(h1_out)
for i in range(config['N_hidden_layers']):
middle_layer = tf.layers.conv1d(
sum_layer,
config['N_latent_variables'],
config['adj_window_size'],
strides=1,
bias_initializer=tf.contrib.layers.xavier_initializer(
uniform=False),
kernel_initializer=tf.contrib.layers.xavier_initializer(
uniform=False),
padding='valid',
activation=None)
middle_layer_norm = tf.layers.batch_normalization(
middle_layer,
axis=2,
training=(mode == learn.ModeKeys.TRAIN)
)
middle_layer_norm = tf.layers.dropout(middle_layer_norm,
rate=config['dropout'],
training=mode == learn.ModeKeys.TRAIN)
middle_layer_out = tf.nn.relu(middle_layer_norm)
summarize_layer('middle_layer%d_out' % i, middle_layer_out)
dw = int((config['adj_window_size'] - 1) / 2)
sum_layer = tf.slice(sum_layer,
[0, dw, 0],
[-1, sum_layer.shape[1].value - 2 * dw, -1])
sum_layer += middle_layer_out
shape = (config['batch_size'], config['window_size'],
config['N_latent_variables'])
flattened_inputs = tf.reshape(sum_layer, [shape[0] * shape[1], shape[2]])
output = tf.layers.dense(flattened_inputs,
units=1,
name='output_layer',
activation=None,
kernel_initializer=tf.random_normal_initializer(
mean=0.0, stddev=0.0001),
bias_initializer=tf.constant_initializer(.2),
)
output = tf.reshape(output, shape=[shape[0], shape[1]])
if config['global_features'] == 'prepost':
post_0 = dense_batch_relu(global_features_processed,
mode == learn.ModeKeys.TRAIN,
'global_post_0',
config['N_global_latent_variables'],
use_relu=True,
weight_scale=.2)
post = dense_batch_relu(post_0,
mode == learn.ModeKeys.TRAIN,
'global_post',
1,
use_relu=False,
weight_scale=.2)
post = tf.log(1 + tf.exp(post))
summarize_layer("post", post)
output = output * post
summarize_layer('cropped_output_before_norm', output)
output = tf.layers.batch_normalization(output,
name='output_normalized',
training=(mode == learn.ModeKeys.TRAIN))
output = tf.nn.relu(output)
summarize_layer('cropped_output', output)
output = tf.reshape(output, shape=(config['batch_size'], config['window_size']))
neuron_ids = tf.identity(features['neuron_ids'])
normalizers = tf.get_variable(
'normalizers_0',
shape=(config['N_neurons'] + 1,),
initializer=tf.ones_initializer(dtype=tf.float32),
trainable=True)
tf.add_to_collection('normalizers_0', tf.GraphKeys.TRAINABLE_VARIABLES)
normalizers = normalizers / (1 + tf.reduce_mean(normalizers))
summarize_layer('normalizers', normalizers)
summary.scalar("fraction_of_zeros_in_output", tf.nn.zero_fraction(output))
outputs = {'output': output}
return outputs, normalizers
def get_config():
config = {'feature_type': 'continuous',
'window_size': 64,
'conv_window_size': 33,
'adj_window_size': 9,
'model_name': "conservative_model",
'eval_every': 2000,
'niter': int(170e3),
'decay_every': int(110e3),
'decay_multiplier': .1,
'mean_obs_per': 3e4,
'num_training_sets': 10,
'batch_size': 128,
'network_fn': residual_network,
'alpha': 1e-4,
'N_neurons': 174,
'N_latent_variables': 32,
'N_global_features': 8, # <= 32, can truncate the representation
'N_global_latent_variables': 4, # the dimensionality of the adaptation
'N_global_intermediate_variables' : 16, # hidden features in the mini-net that does adaptation
'N_hidden_layers': 7,
'validation_cycle': 1,
'validation_bycell': 'sub',
'use_normalizer': True,
'boundary_conditions': 'mirror',
'global_features': 'prepostunsupervised', #'prepost' uses predefined global features instead of the learned ones.
'global_lengthscale': 5000, # for prepost
'dropout': 0.3,
'state_size': 1,
'refine_recording': None,
'early_stopping_rounds': int(1e5),
}
if config['refine_recording'] is not None:
config['eval_every'] /= 5
config['early_stopping_rounds'] = 1e4
config['niter'] = 100000
if config['conv_window_size'] % 2 == 0:
raise NotImplementedError("Even conv_window_size")
window_padding = int(
config['N_hidden_layers'] * (config['adj_window_size'] - 1) / 2 + (config['conv_window_size'] - 1) / 2)
config['window_padding_l'] = window_padding
config['window_padding_r'] = window_padding
config['window_padding'] = window_padding
return config