def atrous_conv_block(self, inputs, depths, stage, rate, is_training, s=2):
depth1, depth2, depth3 = depths
layer1 = relu(
bn(
atrous_conv2d(inputs,
depth1, [1, 1],
rate,
name=stage + '_a_conv'), is_training))
layer2 = relu(
bn(
atrous_conv2d(layer1,
depth2, [3, 3],
rate,
name=stage + '_b_conv'), is_training))
layer3 = bn(
atrous_conv2d(layer2, depth3, [1, 1], rate,
name=stage + '_c_conv'), is_training)
shortcut = bn(
conv2d(inputs,
depth3, [1, 1],
name=stage + '_shortcut',
strides=[1, s, s, 1],
padding='VALID'), is_training)
layer4 = relu(tf.add(layer3, shortcut))
return layer4
def atrous_identity_block(self, inputs, depths, stage, rate, is_training):
depth1, depth2, depth3 = depths
layer1 = relu(
bn(
atrous_conv2d(inputs,
depth1, [1, 1],
rate,
name=stage + '_a_identity'), is_training))
layer2 = relu(
bn(
atrous_conv2d(layer1,
depth2, [3, 3],
rate,
name=stage + '_b_identity'), is_training))
layer3 = bn(
atrous_conv2d(layer2,
depth3, [1, 1],
rate,
name=stage + '_c_identity'), is_training)
layer4 = relu(tf.add(layer3, inputs))
return layer4
def make_model(self, inputs, is_training):
"""
extract feature using ResNet. Encoder
"""
with tf.variable_scope('ResNet50'):
x = conv2d(inputs, 64, [7, 7], strides=[1, 2, 2, 1],
name='conv1') # size 1/2
x = bn(x, is_training)
x = relu(x)
x = max_pool(x, ksize=[1, 3, 3, 1], name='pool1') # size 1/4
x = self.conv_block(x, [64, 64, 256], '2_1', is_training, s=1)
x = self.identity_block(x, [64, 64, 256], '2_2', is_training)
x = self.identity_block(x, [64, 64, 256], '2_3', is_training)
x = self.conv_block(x, [128, 128, 512], '3_1', is_training)
x = self.identity_block(x, [128, 128, 512], '3_2', is_training)
x = self.identity_block(x, [128, 128, 512], '3_3', is_training)
x = self.atrous_conv_block(x, [256, 256, 1024],
'4_1',
2,
is_training,
s=1)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_2', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_3', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_4', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_5', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_6', 2,
is_training)
x = self.atrous_conv_block(x, [512, 512, 2048],
'5_1',
4,
is_training,
s=1)
x = self.atrous_identity_block(x, [512, 512, 2048], '5_2', 4,
is_training)
x = self.atrous_identity_block(x, [512, 512, 2048], '5_3', 4,
is_training)
"""
Astrous Pyrimid Pooling. Decoder
"""
with tf.variable_scope('ASPP'):
rate6 = atrous_conv2d(x, self.N_CLASS, [3, 3], 6, name='rate6')
rate6 = conv2d(rate6, self.N_CLASS, [1, 1], name='rate6_conv1')
rate6 = conv2d(rate6, self.N_CLASS, [1, 1], name='rate6_conv2')
rate12 = atrous_conv2d(x, self.N_CLASS, [3, 3], 12, name='rate12')
rate12 = conv2d(rate12, self.N_CLASS, [1, 1], name='rate12_conv1')
rate12 = conv2d(rate12, self.N_CLASS, [1, 1], name='rate12_conv2')
rate18 = atrous_conv2d(x, self.N_CLASS, [3, 3], 18, name='rate18')
rate18 = conv2d(rate18, self.N_CLASS, [1, 1], name='rate18_conv1')
rate18 = conv2d(rate18, self.N_CLASS, [1, 1], name='rate18_conv2')
rate24 = atrous_conv2d(x, self.N_CLASS, [3, 3], 24, name='rate24')
rate24 = conv2d(rate24, self.N_CLASS, [1, 1], name='rate24_conv1')
rate24 = conv2d(rate24, self.N_CLASS, [1, 1], name='rate24_conv2')
# self.logits = tf.add_n([rate6, rate12, rate18, rate24])
# self.out = tf.image.resize_bilinear(self.logits, size=[192, 192])
add_aspp = tf.add_n([rate6, rate12, rate18, rate24])
logits = tf.image.resize_bilinear(add_aspp,
size=[self.RESIZE, self.RESIZE])
pred = tf.argmax(logits, axis=3)
pred = tf.expand_dims(pred, dim=3)
return logits, pred
def bottleneck(self,
inputs,
out_depth,
f_h,
f_w,
is_training,
keep_prob,
dilated_rate=None,
mode=None,
scope=None):
reduce_depth = int(inputs.get_shape().as_list()[3] / 4)
with tf.variable_scope(scope):
if mode == 'downsampling':
main_branch = max_pool(inputs, name='_pool')
depth_to_pad = abs(inputs.get_shape().as_list()[3] - out_depth)
paddings = tf.convert_to_tensor([[0, 0], [0, 0], [0, 0],
[0, depth_to_pad]])
main_branch = tf.pad(main_branch,
paddings=paddings,
name='_main_padding')
sub_branch = prelu(bn(
conv2d(inputs,
reduce_depth, [2, 2],
name='_conv1',
strides=[1, 2, 2, 1]), is_training),
name='prelu_conv1')
sub_branch = prelu(bn(
conv2d(sub_branch,
reduce_depth, [f_h, f_w],
name='_conv2',
strides=[1, 1, 1, 1]), is_training),
name='prelu_conv2')
sub_branch = prelu(bn(
conv2d(sub_branch,
out_depth, [1, 1],
name='_conv3',
strides=[1, 1, 1, 1]), is_training),
name='prelu_conv3')
sub_branch = prelu(spatial_dropout(sub_branch, keep_prob),
name='prelu_dropout')
out = prelu(tf.add(main_branch, sub_branch), name='prelu_add')
return out
elif mode == 'dilated':
main_branch = inputs
sub_branch = prelu(bn(
conv2d(
inputs,
reduce_depth,
[1, 1],
name='_conv1',
), is_training),
name='prelu_conv1')
sub_branch = prelu(bn(
atrous_conv2d(sub_branch,
reduce_depth, [f_h, f_w],
dilated_rate,
name='_conv2'), is_training),
name='prelu_conv2')
sub_branch = prelu(bn(
conv2d(inputs, out_depth, [1, 1], name='_conv3'),
is_training),
name='prelu_conv3')
sub_branch = prelu(spatial_dropout(sub_branch, keep_prob),
name='prelu_dropout')
out = prelu(tf.add(main_branch, sub_branch), name='prelu_add')
return out
elif mode == 'asymmetric':
main_branch = inputs
sub_branch = prelu(bn(
conv2d(inputs, reduce_depth, [1, 1], name='_conv1'),
is_training),
name='prelu_conv1')
sub_branch = prelu(bn(
conv2d(sub_branch, reduce_depth, [f_h, 1], name='_conv2'),
is_training),
name='prelu_conv2')
sub_branch = prelu(bn(
conv2d(sub_branch, reduce_depth, [1, f_w], name='_conv3'),
is_training),
name='prelu_conv3')
sub_branch = prelu(bn(
conv2d(sub_branch, out_depth, [1, 1], name='_conv4'),
is_training),
name='prelu_conv4')
sub_branch = prelu(spatial_dropout(sub_branch, keep_prob),
name='prelu_dropout')
out = prelu(tf.add(main_branch, sub_branch), name='prelu_add')
return out
elif mode == 'upsampling':
# 논문에서 나오는 unpool 대신 bilinear interpolation 사용
in_shape = inputs.get_shape().as_list()
main_branch = tf.image.resize_bilinear(
inputs, size=[in_shape[1] * 2, in_shape[2] * 2])
main_branch = prelu(bn(
conv2d(main_branch, out_depth, [3, 3], name='_conv0'),
is_training),
name='prelu_conv1')
sub_branch = prelu(bn(
conv2d(inputs, reduce_depth, [1, 1], name='_conv1'),
is_training),
name='prelu_conv2')
sub_branch = prelu(bn(
conv2d_t(sub_branch, [
in_shape[0], in_shape[1] * 2, in_shape[2] * 2,
reduce_depth
], [3, 3],
name='_conv2'), is_training),
name='prelu_conv3')
sub_branch = prelu(bn(
conv2d(sub_branch, out_depth, [1, 1], name='_conv3'),
is_training),
name='prelu_conv4')
sub_branch = prelu(spatial_dropout(sub_branch, keep_prob),
name='prelu_dropout')
out = prelu(tf.add(main_branch, sub_branch), name='prelu_add')
return out
elif mode == 'normal':
main_branch = inputs
sub_branch = prelu(bn(
conv2d(inputs,
reduce_depth, [1, 1],
name='_conv1',
strides=[1, 1, 1, 1]), is_training),
name='prelu_conv1')
sub_branch = prelu(bn(
conv2d(sub_branch,
reduce_depth, [f_h, f_w],
name='_conv2',
strides=[1, 1, 1, 1]), is_training),
name='prelu_conv2')
sub_branch = prelu(bn(
conv2d(sub_branch,
out_depth, [1, 1],
name='_conv3',
strides=[1, 1, 1, 1]), is_training),
name='prelu_conv3')
sub_branch = prelu(spatial_dropout(sub_branch, keep_prob),
name='prelu_dropout')
out = prelu(tf.add(main_branch, sub_branch), name='prelu_add')
return out
def make_model(self, inputs, is_training):
"""
extract feature using ResNet. Encoder
"""
with tf.variable_scope('ResNet50'):
x = conv2d(inputs, 64, [7, 7], strides=[1, 2, 2, 1],
name='conv1') # size 1/2
x = bn(x, is_training)
x = relu(x)
x = max_pool(x, ksize=[1, 3, 3, 1], name='pool1') # size 1/4
x = self.conv_block(x, [64, 64, 256], '2_1', is_training, s=1)
x = self.identity_block(x, [64, 64, 256], '2_2', is_training)
x = self.identity_block(x, [64, 64, 256], '2_3', is_training)
x = self.conv_block(x, [128, 128, 512], '3_1', is_training)
x = self.identity_block(x, [128, 128, 512], '3_2', is_training)
x = self.identity_block(x, [128, 128, 512], '3_3', is_training)
x = self.atrous_conv_block(x, [256, 256, 1024],
'4_1',
2,
is_training,
s=1)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_2', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_3', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_4', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_5', 2,
is_training)
x = self.atrous_identity_block(x, [256, 256, 1024], '4_6', 2,
is_training)
x = self.atrous_conv_block(x, [512, 512, 2048],
'5_1',
4,
is_training,
s=1)
x = self.atrous_identity_block(x, [512, 512, 2048], '5_2', 4,
is_training)
x = self.atrous_identity_block(x, [512, 512, 2048], '5_3', 4,
is_training)
"""
Astrous Pyrimid Pooling. Decoder
"""
with tf.variable_scope('ASPP'):
feature_map_shape = x.get_shape().as_list()
# global average pooling
# feature 맵의 height, width를 평균을 낸다.
feature_map = tf.reduce_mean(x, [1, 2], keepdims=True)
feature_map = conv2d(feature_map,
256, [1, 1],
name='gap_feature_map')
feature_map = tf.image.resize_bilinear(
feature_map, [feature_map_shape[1], feature_map_shape[2]])
rate1 = conv2d(x, 256, [1, 1], name='rate1')
rate6 = atrous_conv2d(x, 256, [3, 3], rate=6, name='rate6')
rate12 = atrous_conv2d(x, 256, [3, 3], rate=12, name='rate12')
rate18 = atrous_conv2d(x, 256, [3, 3], rate=18, name='rate18')
concated = tf.concat([feature_map, rate1, rate6, rate12, rate18],
axis=3)
net = conv2d(concated, 256, [1, 1], name='net')
logits = conv2d(net, self.N_CLASS, [1, 1], name='logits')
logits = tf.image.resize_bilinear(logits,
size=[self.RESIZE, self.RESIZE],
name='out')
pred = tf.argmax(logits, axis=3)
pred = tf.expand_dims(pred, dim=3)
return logits, pred