def first_residual_atrous_block(x,
kernel,
out_channel,
strides,
is_train,
name="unit"):
input_channels = x.get_shape().as_list()[-1]
with tf.variable_scope(name) as scope:
print('\tBuilding residual unit: %s' % scope.name)
# Shortcut connection
if input_channels == out_channel:
if strides == 1:
shortcut = tf.identity(
x
) # returns a tensor with the same shape and contents as x
else:
shortcut = tf.nn.max_pool(x, [1, 1, 1, 1], [1, 1, 1, 1],
'VALID')
else:
in_shape = x.get_shape()
shortcut = atrous_conv_layer(
x, [1, 1, in_shape[-1], out_channel],
out_channel,
strides,
name='shortcut') # 1x1 conv to obtain out_channel maps
# Residual
in_shape = x.get_shape()
x = atrous_conv_layer(x, [kernel, kernel, in_shape[-1], out_channel],
out_channel,
strides,
name='conv_1')
x = mygn(x, name='gn_1')
x = myrelu(x, name='relu_1')
in_shape = x.get_shape()
x = atrous_conv_layer(x, [kernel, kernel, in_shape[-1], out_channel],
out_channel,
strides,
name='conv_2')
x = mygn(x, name='gn_2')
# Merge
x = x + shortcut
x = myrelu(x, name='relu_2')
return x
def atrous_spatial_pyramid_pooling_block(x, is_train, depth=256, name='aspp'):
in_shape = x.get_shape()
input_size = tf.shape(x)[1:3]
filters = [1, 4, 3, 3, 1, 1]
atrous_rates = [1, 6, 12, 18, 1, 1]
with tf.variable_scope(name) as scope:
print('\tBuilding aspp unit: %s' % scope.name)
# Branch 0: 1x1 conv
branch0 = conv_layer(x, [filters[0], filters[0], in_shape[3], depth],
name="branch0")
branch0 = mygn(branch0, name='gn_0')
# Branch 1: 3x3 atrous_conv (rate = 6)
branch1 = atrous_conv_layer(
x, [filters[1], filters[1], in_shape[-1], depth],
depth,
atrous_rates[1],
name='branch1')
branch1 = mygn(branch1, name='gn_1')
# Branch 2: 3x3 atrous_conv (rate = 12)
branch2 = atrous_conv_layer(
x, [filters[2], filters[2], in_shape[-1], depth],
depth,
atrous_rates[2],
name='branch2')
branch2 = mygn(branch2, name='gn_2')
# Branch 3: 3x3 atrous_conv (rate = 18)
branch3 = atrous_conv_layer(
x, [filters[3], filters[3], in_shape[-1], depth],
depth,
atrous_rates[3],
name='branch3')
branch3 = mygn(branch3, name='gn_3')
# Branch 4: image pooling
# 4.1 global average pooling
branch4 = tf.reduce_mean(x, [1, 2],
name='global_average_pooling',
keepdims=True)
# 4.2 1x1 convolution with 256 filters and batch normalization
branch4 = conv_layer(x, [filters[4], filters[4], in_shape[3], depth],
name="brach4")
branch4 = mygn(branch4, name='gn_4')
# 4.3 bilinearly upsample features
branch4 = tf.image.resize_bilinear(branch4,
input_size,
name='branch4_upsample')
# Output
out = tf.concat([branch0, branch1, branch2, branch3, branch4],
axis=3,
name='aspp_concat')
out = myrelu(out, name='relu_out')
in_shape = out.get_shape()
out = conv_layer(out, [filters[5], filters[5], in_shape[3], depth],
name="aspp_out",
relu='no')
return out
def residual_block(x, kernel, is_train, name="unit"):
num_channel = x.get_shape().as_list()[-1]
with tf.variable_scope(name) as scope:
print('\tBuilding residual unit: %s' % scope.name)
# Shortcut connection
shortcut = x
# Residual
in_shape = x.get_shape()
x = conv_layer(x, [kernel, kernel, in_shape[3], num_channel],
strides=1,
name="conv_1")
x = mygn(x, name='gn_1')
x = myrelu(x, name='relu_1')
in_shape = x.get_shape()
conv_layer(x, [kernel, kernel, in_shape[3], num_channel],
strides=1,
name="conv_2")
x = mygn(x, name='gn_2')
# Merge
x = x + shortcut
x = myrelu(x, name='relu_2')
return x
def deeplab_net(X, classes, is_train=tf.constant(True), pkeep=1): # ResNet 18
if pkeep != 1:
print('\nBuilding train model')
else:
print('\nBuilding validation model')
# filters = [128, 128, 256, 512, 1024]
filters = [64, 64, 128, 256, 512]
kernels = [7, 3, 3, 3, 3]
strides = [2, 0, 2, 2, 2]
size = tf.shape(X)
height = size[1]
width = size[2]
with tf.variable_scope('generator'):
# conv1
print('\tBuilding unit: conv1')
with tf.variable_scope('conv1'):
in_shape = X.get_shape()
x = conv_layer(X,
[kernels[0], kernels[0], in_shape[3], filters[0]],
name="conv")
x = mygn(x, name="gn")
x = myrelu(x)
x = tf.nn.max_pool(x, [1, 3, 3, 1], [1, 2, 2, 1], 'SAME')
# conv2_x
x = residual_block(x, kernels[1], is_train, name='conv2_1')
x = residual_block(x, kernels[1], is_train, name='conv2_2')
# conv3_x
x = first_residual_block(x,
kernels[2],
filters[2],
strides[2],
is_train,
name='conv3_1')
x = residual_block(x, kernels[2], is_train, name='conv3_2')
# conv4_x
x = first_residual_block(x,
kernels[3],
filters[3],
strides[3],
is_train,
name='conv4_1')
x = residual_block(x, kernels[3], is_train, name='conv4_2')
# conv5_x
x = first_residual_atrous_block(x,
kernels[4],
filters[4],
strides[4],
is_train,
name='conv5_1')
x = residual_atrous_block(x, kernels[4], is_train, name='conv5_2')
# aspp
x = atrous_spatial_pyramid_pooling_block(x,
is_train,
depth=256,
name='aspp_1')
print('\tBuilding unit: class scores') # Maybe another layer ???
in_shape = x.get_shape()
x = conv_layer(x, [1, 1, in_shape[3], classes], name="class_scores")
# upsample logits
print('\tBuilding unit: upsample')
logits = tf.image.resize_images(x,
tf.stack([height, width]),
method=tf.image.ResizeMethod.BILINEAR,
align_corners=False)
return logits
#X = tf.ones([1, 256, 256, 3], dtype=tf.float32)
#build_model(X, is_train=tf.constant(True), pkeep=1)