/
shufflenet.py
148 lines (136 loc) · 8.94 KB
/
shufflenet.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
import tensorflow as tf
import numpy as np
import time
SHUFFLENET_MEAN = [103.939, 116.779, 123.68]
NORMALIZER = 0.017
#Parameters credits to https://github.com/MG2033/ShuffleNet
class Shufflenet:
#Load pretrained model on initialization. Model downloaded from http://models.tensorpack.com/ImageNetModels/ShuffleNetV1-1x-g=8.npz
def __init__(self, model_loc):
self.trained_model = np.load(model_loc, encoding = 'latin1')
print("Pre-trained npz model loaded")
# Uncomment below 2 lines to check model entries (Kernels for conv), (mean, variance, beta and gamma for BN) and (weights and biases) for final FC layer
for x in self.trained_model.files:
print(x + " " + str(self.trained_model[x].shape))
'''
Point-wise group convolution operation.
Inputs: Activations of shape [N, H, W, C], Stage in format
'stagex', block in format 'blockx' and layer in
format 'convx', num_groups are the number of
groups to split the activations and kernels in.
Outputs: Output activations of group convolution result
'''
def pw_gconv(self, activations, stage, block, layer, num_groups, name):
with tf.name_scope(name):
layer_name = str(stage) + '/' + str(block) + '/' + str(layer) + '/W:0'
kernels = self.trained_model[layer_name]
print(kernels.shape)
ch_per_group = activations.shape[3] // num_groups
act_split = tf.split(activations, num_or_size_splits = num_groups, axis = 3)
kernels_split = tf.split(kernels, num_or_size_splits = num_groups, axis = 3)
convs = []
for grp in range(0, num_groups):
convs.append(tf.nn.conv2d(act_split[grp], kernels_split[grp], padding = 'SAME', strides = [1, 1, 1, 1], data_format = 'NHWC', name='pw_gconv_' + str(grp)))
return tf.concat(convs, axis = 3)
'''
Depth-wise convolution operation.
Inputs: Activations of shape [N, H, W, C], stage in format 'stagex',
block in format 'blockx', padding, stride and name to give to
the node in tensorboard visualization
Outputs: Output activations of the dw conv operation
'''
def dw_conv(self, activations, stage, block, padding = 'SAME', stride = 1, name="dw_conv"):
with tf.name_scope(name):
inp_ch = activations.shape[3]
act_shape = activations.shape
layer_name = str(stage) + '/' + str(block) + '/dconv/W:0'
kernels = self.trained_model[layer_name]
print(kernels.shape)
kernel_size = kernels.shape[0]
conv_result = tf.nn.depthwise_conv2d(activations, kernels, [1, stride, stride, 1], padding = padding, data_format = 'NHWC', name='dw_conv_' + stage + '_' + block)
return conv_result
'''
Batch Normalization operations
Inputs: Activations of shape [N, H, W, C], stage in format 'stagex',
block in format 'blockx', layer in format 'convx' and name to
give to the node in tensorboard graph summary.
Outputs: Output activations of BN operation
'''
def batch_normalization(self, activations, stage, block, layer, name):
with tf.name_scope(name):
layer_name = str(stage) + '/' + str(block) + '/' if stage is not '' else ''
layer_name = layer_name + 'conv1/bn/' if layer == 'conv1' else layer_name + layer+'_bn/'
bn_out = tf.nn.batch_normalization(activations, self.trained_model[layer_name + 'mean/EMA:0'], self.trained_model[layer_name + 'variance/EMA:0'], self.trained_model[layer_name + 'beta:0'], self.trained_model[layer_name + 'gamma:0'], variance_epsilon=0.001, name = 'bn_' + stage + '_' + block + '_' + layer if stage is not '' else 'bn_conv1')
return bn_out
'''
Channel Shuffle Operation. (Credits to Tensorpack Shufflenet Implementation https://github.com/tensorpack/tensorpack/blob/master/examples/ImageNetModels/shufflenet.py)
Inputs: Activations of shape [N, H, W, C], num_groups = 8,
Name to give to the node in tensorboard graph summary.
Outputs Activations after ch shuffle op.
'''
def channel_shuffle(self, activations, num_groups = 8, name='ch_shuffle'):
with tf.name_scope(name):
activations = tf.transpose(activations, perm = [0, 3, 1, 2])
in_shape = activations.get_shape().as_list()
in_channel = in_shape[1]
l = tf.reshape(activations, [-1, in_channel // num_groups, num_groups] + in_shape[-2:])
l = tf.transpose(l, [0, 2, 1, 3, 4])
l = tf.reshape(l, [-1, in_channel] + in_shape[-2:])
l = tf.transpose(l, perm = [0, 2, 3, 1])
return l
def shufflenet_unit(self, activations, stage, block, stride, num_groups=8, name="shufflenet_unit"):
with tf.name_scope(name):
residual = activations
num_split = num_groups if activations.shape[3] > 24 else 1
print("Stage: " + stage + " block: " + block)
print("inp_act size: " + str(activations.shape))
pwgconv1 = self.pw_gconv(activations, stage, block, 'conv1', num_split, name= stage + "_" + block + "_pwgconv1")
print("pwgconv1 size: " + str(pwgconv1.shape))
bnconv1 = self.batch_normalization(pwgconv1, stage, block, 'conv1', name = stage + "_" + block + "_pwgconv1_batch_norm")
reluconv1 = tf.nn.relu(bnconv1)
ch_sh = self.channel_shuffle(reluconv1, num_groups, name = stage + '_' + block + '_ch_shuffle')
dconv = self.dw_conv(ch_sh, stage, block, padding = 'SAME', stride = stride, name = stage + "_" + block + "_dwconv")
print("dconv size: " + str(dconv.shape))
bndconv = self.batch_normalization(dconv, stage, block, 'dconv', name = stage + "_" + block + "_dconv_batch_norm")
pwgconv2 = self.pw_gconv(bndconv, stage, block, 'conv2', num_groups, name= stage + "_" + block + "_pwgconv2")
print("pwgconv2 size: " + str(pwgconv2.shape))
bnconv2 = self.batch_normalization(pwgconv2, stage, block, 'conv2', name = stage + "_" + block + "_pwgconv2_batch_norm")
if stride == 1:
return tf.nn.relu(bnconv2 + residual, name = 'relu_' + stage + '_' + block)
elif stride == 2:
residual = tf.nn.avg_pool(residual, [1, 3, 3, 1], strides = [1, 2, 2, 1], padding = 'SAME', data_format = 'NHWC', name = 'avg_pool_' + stage + '_' + block)
return tf.concat([residual, tf.nn.relu(bnconv2)], axis = 3, name = 'concat_' + stage + '_' + block)
else:
raise ValueError("Stride value can only be 1 or 2 for Shufflenet")
def shufflenet_stage(self, activations, stage, repeat, num_groups=8, name = "shufflenet_stage"):
with tf.name_scope(name):
first_block = self.shufflenet_unit(activations, stage, 'block0', stride = 2, num_groups = 8, name = "shufflenet_unit_" + stage + "_block0")
res = first_block
for b in range(1, repeat+1):
res = self.shufflenet_unit(res, stage, 'block' + str(b), stride = 1, num_groups = 8, name = "shufflenet_unit_" + stage + "_block" + str(b))
return res
def shufflenet_stage1(self, activations):
with tf.name_scope("shufflenet_stage1"):
kernels = self.trained_model['conv1/W:0']
res = tf.nn.conv2d(activations, kernels, padding = 'SAME', strides = [1, 2, 2, 1], data_format = 'NHWC', name = 'Conv1')
res = self.batch_normalization(res, '', '', 'conv1', name = 'stage1_conv2d_batch_norm')
res = tf.nn.max_pool(res, [1, 3, 3, 1], strides = [1, 2, 2, 1], padding = 'SAME', data_format = 'NHWC', name = 'MaxPool1')
return res
def fc_layer(self, activations):
with tf.name_scope('fc_layer'):
layer_name = 'linear'
weights = self.trained_model[layer_name + '/W:0']
biases = self.trained_model[layer_name + '/b:0']
flattened_out = tf.contrib.layers.flatten(activations)
return tf.nn.bias_add(tf.matmul(flattened_out, weights), biases)
def build(self, image):
red, green, blue = tf.split(axis=3, num_or_size_splits=3, value=image)
bgr = tf.concat(axis=3, values=[(blue - SHUFFLENET_MEAN[0])*NORMALIZER, (green - SHUFFLENET_MEAN[1])*NORMALIZER, (red - SHUFFLENET_MEAN[2])*NORMALIZER])
stage1 = self.shufflenet_stage1(bgr)
stage2 = self.shufflenet_stage(stage1, 'stage2', repeat = 3, num_groups = 8, name = "shufflenet_stage2")
stage3 = self.shufflenet_stage(stage2, 'stage3', repeat = 7, num_groups = 8, name = "shufflenet_stage3")
stage4 = self.shufflenet_stage(stage3, 'stage4', repeat = 3, num_groups = 8, name = "shufflenet_stage4")
g_pool = tf.nn.avg_pool(stage4, [1, 7, 7, 1], strides = [1, 1, 1, 1], padding = 'VALID', data_format = 'NHWC', name = 'GlobalPool')
logits = self.fc_layer(g_pool)
logits = tf.nn.softmax(logits, name = "SoftMax_unit")
return logits