-
Notifications
You must be signed in to change notification settings - Fork 5
/
converter.py
194 lines (179 loc) · 9.22 KB
/
converter.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
import tensorflow as tf
import numpy as np
class Converter(object):
def __init__(self, tf_nodes, mx_nodes, mx_params):
self.tf_nodes = tf_nodes
self.mx_nodes = mx_nodes
self.mx_params = mx_params
def to_tuple(self, string, conv_type=str):
return tuple(map(conv_type, map(str.strip, string[1:-1].split(','))))
def create_var(self, node, shape=None):
node_name = node['name']
if shape is None:
if node_name in self.mx_params:
shape = self.mx_params[node_name].shape
else:
shape = ()
# print('Creating var with shape:', shape)
created_node = tf.get_variable(node_name, shape=shape, initializer=tf.zeros_initializer)
self.tf_nodes[node_name] = created_node
# if node_name in params:
# tf_nodes[node_name].load(params[node_name].asnumpy())
return created_node
def create_bn(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
epsilon = float(node['attr']['eps'])
input_shape = input_sym.get_shape()
axis = list(range(len(input_shape) - 1))
def create_bn_params(i):
cur_node = self.mx_nodes[node['inputs'][i][0]]
cur_name = cur_node['name']
self.create_var(cur_node)
self.tf_nodes[cur_name].load(self.mx_params[cur_name].asnumpy())
return self.tf_nodes[cur_name]
if len(node['inputs']) > 3:
gamma, beta, mean, var = (create_bn_params(i) for i in range(1, 5))
else:
gamma, beta = (create_bn_params(i) for i in range(1, 3))
mean = tf.get_variable(node_name + '_mean', shape=input_shape[-1], initializer=tf.zeros_initializer)
mean.load(np.zeros((input_shape[-1],), dtype='float32'))
var = tf.get_variable(node_name + '_var', shape=input_shape[-1], initializer=tf.ones_initializer)
var.load(np.ones((input_shape[-1],), dtype='float32'))
# TODO: add support for swtiching between train and inference phases
# For inference use_global_stats=False is ignored
#
# if 'use_global_stats' in node['attr']:
# if node['attr']['use_global_stats'] == 'False':
# # print('Not use')
# mean, var = tf.nn.moments(input_sym, axis)
# else:
# mean, var = tf.nn.moments(input_sym, axis)
if 'fix_gamma' in node['attr']:
if node['attr']['fix_gamma'] == 'True':
# print('Fix')
gamma = tf.get_variable(node_name + '_gamma_fixed', shape=input_shape[-1], initializer=tf.ones_initializer)
gamma.load(np.ones((input_shape[-1],), dtype='float32'))
else:
gamma = tf.get_variable(node_name + '_gamma_fixed', shape=input_shape[-1], initializer=tf.ones_initializer)
gamma.load(np.ones((input_shape[-1],), dtype='float32'))
self.tf_nodes[node_name] = tf.nn.batch_normalization(input_sym, mean, var, beta, gamma, epsilon, name=node_name)
return self.tf_nodes[node_name]
def create_conv(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
num_filters_in = input_sym.get_shape()[-1]
num_filters_out = int(node['attr']['num_filter'])
kernel_size = self.to_tuple(node['attr']['kernel'], int)
# TODO: add bias support
# add_bias = node['attr']['no_bias'] != 'True'
if 'num_group' in node['attr']:
num_group = int(node['attr']['num_group'])
else:
num_group = 1
if 'pad' in node['attr']:
padding = self.to_tuple(node['attr']['pad'], int)
else:
padding = (0, 0)
stride = self.to_tuple(node['attr']['stride'], int)
weights_node = self.mx_nodes[node['inputs'][1][0]]
weights = self.create_var(weights_node,
shape=(kernel_size[0], kernel_size[1], num_filters_in // num_group, num_filters_out))
weights_numpy = self.mx_params[weights_node['name']].asnumpy().transpose((2, 3, 1, 0))
if padding[0] > 0 or padding[1] > 0:
padded_input = tf.pad(input_sym, [[0, 0], [padding[0], padding[0]], [padding[1], padding[1]], [0, 0]], 'CONSTANT')
else:
padded_input = input_sym
convolve = lambda input_sym, kernel, name=None: tf.nn.conv2d(input_sym, kernel, [1, stride[0], stride[1], 1], padding='VALID', name=name)
weights.load(weights_numpy)
if num_group > 1:
input_groups = tf.split(axis=3, num_or_size_splits=num_group, value=padded_input)
weight_groups = tf.split(axis=3, num_or_size_splits=num_group, value=weights)
output_groups = [convolve(i, k) for i, k in zip(input_groups, weight_groups)]
self.tf_nodes[node_name] = tf.concat(axis=3, values=output_groups, name=node_name)
else:
self.tf_nodes[node_name] = convolve(padded_input, weights, name=node_name)
return self.tf_nodes[node_name]
def create_pooling(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
pooling_type = node['attr']['pool_type']
kernel_size = self.to_tuple(node['attr']['kernel'], int)
if 'stride' in node['attr']:
stride = self.to_tuple(node['attr']['stride'], int)
else:
stride = (1, 1)
if 'global_pool' in node['attr']:
global_pool = node['attr']['global_pool'] == 'True'
else:
global_pool = False
if 'pad' in node['attr']:
padding = self.to_tuple(node['attr']['pad'], int)
else:
padding = (0, 0)
if global_pool:
self.tf_nodes[node_name] = tf.reduce_mean(input_sym, reduction_indices=[1, 2], name=node_name)
else:
if padding[0] > 0 or padding[1] > 0:
padded_input = tf.pad(input_sym,
[[0, 0], [padding[0], padding[0]], [padding[1], padding[1]], [0, 0]],
'CONSTANT')
else:
padded_input = input_sym
if pooling_type == 'max':
self.tf_nodes[node_name] = tf.nn.max_pool(padded_input,
ksize=[1, kernel_size[0], kernel_size[1], 1],
strides=[1, stride[0], stride[1], 1],
padding='VALID', name=node_name)
else:
raise NameError('Unknown pooling type: %s' % pooling_type)
return self.tf_nodes[node_name]
def create_activation(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
activation_type = node['attr']['act_type']
# TODO: more activation types
if activation_type == 'relu':
activation_fn = tf.nn.relu
else:
raise NameError('Unknown activation type: %s' % activation_type)
self.tf_nodes[node_name] = activation_fn(input_sym, name=node_name)
return self.tf_nodes[node_name]
def create_softmax(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
self.tf_nodes[node_name] = tf.nn.softmax(input_sym, name=node_name)
return self.tf_nodes[node_name]
def create_elementwise(self, node, op='sum'):
node_name = node['name']
inputs_sym = [self.tf_nodes[self.mx_nodes[n[0]]['name']] for n in node['inputs']]
# TODO: more elementwise types
if op == 'sum':
self.tf_nodes[node_name] = tf.add_n(inputs_sym, name=node_name)
else:
raise NameError('Unknown elementwise type: %s' % op)
return self.tf_nodes[node_name]
def create_fc(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
num_units_in = input_sym.get_shape()[1]
num_units_out = int(node['attr']['num_hidden'])
weights_node = self.mx_nodes[node['inputs'][1][0]]
weights = self.create_var(weights_node, shape=(num_units_in, num_units_out))
bias_node = self.mx_nodes[node['inputs'][2][0]]
bias = self.create_var(bias_node, shape=(num_units_out,))
weights_numpy = self.mx_params[weights_node['name']].asnumpy()
weights.load(weights_numpy.T)
bias.load(self.mx_params[bias_node['name']].asnumpy())
self.tf_nodes[node_name] = tf.nn.xw_plus_b(input_sym, weights, bias, name=node_name)
return self.tf_nodes[node_name]
def create_norm(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
self.tf_nodes[node_name] = tf.nn.l2_normalize(input_sym, dim=1, name=node_name)
return self.tf_nodes[node_name]
def create_flatten(self, node):
node_name = node['name']
input_sym = self.tf_nodes[self.mx_nodes[node['inputs'][0][0]]['name']]
self.tf_nodes[node_name] = tf.contrib.layers.flatten(input_sym)
return self.tf_nodes[node_name]