def add_deconv(self, bilinear=False):
fusion = self.get_output('fusion')
with tf.variable_scope('deconv') as scope:
# Learn from scratch
if not bilinear:
w_deconv = tf.get_variable(
'weights', [16, 16, self.num_classes, self.num_classes],
initializer=tf.truncated_normal_initializer(0.0,
stddev=0.01))
# Using fiexed bilinearing upsampling filter
else:
w_deconv = tf.get_variable(
'weights',
trainable=True,
initializer=bilinear_upsample_weights(
16, self.num_classes))
b_deconv = tf.get_variable('biases', [self.num_classes],
initializer=tf.constant_initializer(0))
z_deconv = tf.nn.conv2d_transpose(
fusion,
w_deconv, [
self.batch_num, self.max_size[0], self.max_size[1],
self.num_classes
],
strides=[1, 8, 8, 1],
padding='SAME',
name='z') + b_deconv
# Add to store dicts
self.outputs['deconv'] = z_deconv
self.layers['deconv'] = {'weights': w_deconv, 'biases': b_deconv}
def add_shortcut(self, bilinear=True):
conv8 = self.get_output('conv8')
pool4 = self.get_output('pool4')
target_size = int(pool4.get_shape()[1])
with tf.variable_scope('2x_conv8') as scope:
# Learn from scratch
if not bilinear:
w_deconv = tf.get_variable(
'weights', [4, 4, self.num_classes, self.num_classes],
initializer=tf.truncated_normal_initializer(0.0,
stddev=0.01))
# Using fiexed bilinearing upsampling filter
else:
w_deconv = tf.get_variable(
'weights',
trainable=True,
initializer=bilinear_upsample_weights(2, self.num_classes))
b_deconv = tf.get_variable('biases', [self.num_classes],
initializer=tf.constant_initializer(0))
z_deconv = tf.nn.conv2d_transpose(
conv8,
w_deconv,
[self.batch_num, target_size, target_size, self.num_classes],
strides=[1, 2, 2, 1],
padding='SAME',
name='z') + b_deconv
with tf.variable_scope('pool4_1x1') as scope:
w_pool4 = tf.get_variable(
'weights', [1, 1, 512, self.num_classes],
initializer=tf.truncated_normal_initializer(0.0, stddev=0.01))
b_pool4 = tf.get_variable('biases', [self.num_classes],
initializer=tf.constant_initializer(0))
z_pool4 = tf.nn.conv2d(
pool4, w_pool4, strides=[1, 1, 1, 1], padding='SAME') + b_pool4
# Element-wise sum
fusion = z_deconv + z_pool4
#pdb.set_trace()
# Add to store dicts
self.outputs['2x_conv8'] = z_deconv
self.outputs['pool4_1x1'] = z_pool4
self.outputs['fusion'] = fusion
self.layers['2x_conv8'] = {'weights': w_deconv, 'biases': b_deconv}
self.layers['pool4_1x1'] = {'weights': w_pool4, 'biases': b_pool4}
def add_deconv(self, bilinear=False):
fusion = self.get_output('fusion')
with tf.variable_scope('deconv') as scope:
# Learn from scratch
if not bilinear:
w_deconv = tf.get_variable(
'weights', [32, 32, self.num_classes, self.num_classes],
initializer=tf.truncated_normal_initializer(0.0,
stddev=0.01))
# Using fiexed bilinearing upsampling filter
else:
w_deconv = tf.get_variable(
'weights',
trainable=True,
initializer=bilinear_upsample_weights(
16, self.num_classes))
b_deconv = tf.get_variable('biases', [self.num_classes],
initializer=tf.constant_initializer(0))
z_deconv = tf.nn.conv2d_transpose(
fusion,
w_deconv, [
self.batch_num, self.max_size[0], self.max_size[1],
self.num_classes
],
strides=[1, 16, 16, 1],
padding='SAME',
name='z') + b_deconv
'''
try to add crfRnn layer,you can also uncomment it
'''
crf_deconv = CrfRnnLayer(
image_dims=(self.max_size[0], self.max_size[1]),
num_classes=self.num_classes,
theta_alpha=160.,
theta_beta=3.,
theta_gamma=3.,
num_iterations=10,
name='crfrnn')([z_deconv, self.img]) #shape=(1, 640, 640, 21)
# Add to store dicts
self.outputs[
'deconv'] = crf_deconv #<shape=(5, 640, 640, 21) dtype=float32>
#self.outputs['deconv'] = z_deconv#<shape=(5, 640, 640, 21) dtype=float32>
self.layers['deconv'] = {'weights': w_deconv, 'biases': b_deconv}
def add_deconv(self):
conv5_3 = self.get_output('conv5_3')
conv4_3 = self.get_output('conv4_3')
conv3_3 = self.get_output('conv3_3')
conv2_2 = self.get_output('conv2_2')
conv1_2 = self.get_output('conv1_2')
shape1 = tf.shape(conv1_2)[1]
shape2 = tf.shape(conv1_2)[2]
# ======== Preprocessing to align feature dim ======== #
with tf.variable_scope('align5') as scope:
w_conv5 = tf.get_variable(
'weights', [1, 1, 512, 64],
initializer=tf.truncated_normal_initializer(0.0, stddev=0.01))
b_conv5 = tf.get_variable('biases', [64],
initializer=tf.constant_initializer(0))
z_conv5 = tf.nn.conv2d(
conv5_3, w_conv5, strides=[1, 1, 1, 1
], padding='SAME') + b_conv5
with tf.variable_scope('align4') as scope:
w_conv4 = tf.get_variable(
'weights', [1, 1, 512, 64],
initializer=tf.truncated_normal_initializer(0.0, stddev=0.01))
b_conv4 = tf.get_variable('biases', [64],
initializer=tf.constant_initializer(0))
z_conv4 = tf.nn.conv2d(
conv4_3, w_conv4, strides=[1, 1, 1, 1
], padding='SAME') + b_conv4
with tf.variable_scope('align3') as scope:
w_conv3 = tf.get_variable(
'weights', [1, 1, 256, 64],
initializer=tf.truncated_normal_initializer(0.0, stddev=0.01))
b_conv3 = tf.get_variable('biases', [64],
initializer=tf.constant_initializer(0))
z_conv3 = tf.nn.conv2d(
conv3_3, w_conv3, strides=[1, 1, 1, 1
], padding='SAME') + b_conv3
with tf.variable_scope('align2') as scope:
w_conv2 = tf.get_variable(
'weights', [1, 1, 128, 64],
initializer=tf.truncated_normal_initializer(0.0, stddev=0.01))
b_conv2 = tf.get_variable('biases', [64],
initializer=tf.constant_initializer(0))
z_conv2 = tf.nn.conv2d(
conv2_2, w_conv2, strides=[1, 1, 1, 1
], padding='SAME') + b_conv2
# ======== Upsample and skip connection ======== #
with tf.variable_scope('deconv5') as scope:
# Using fiexed bilinearing upsampling filter
w_deconv5 = tf.get_variable('weights',
trainable=False,
initializer=bilinear_upsample_weights(
32, 64))
z_deconv5 = tf.nn.conv2d_transpose(
z_conv5,
w_deconv5, [self.batch_num, shape1, shape2, 64],
strides=[1, 16, 16, 1],
padding='SAME',
name='z')
with tf.variable_scope('deconv4') as scope:
# Using fiexed bilinearing upsampling filter
w_deconv4 = tf.get_variable('weights',
trainable=False,
initializer=bilinear_upsample_weights(
16, 64))
z_deconv4 = tf.nn.conv2d_transpose(
z_conv4,
w_deconv4, [self.batch_num, shape1, shape2, 64],
strides=[1, 8, 8, 1],
padding='SAME',
name='z')
with tf.variable_scope('deconv3') as scope:
# Using fiexed bilinearing upsampling filter
w_deconv3 = tf.get_variable('weights',
trainable=False,
initializer=bilinear_upsample_weights(
8, 64))
z_deconv3 = tf.nn.conv2d_transpose(
z_conv3,
w_deconv3, [self.batch_num, shape1, shape2, 64],
strides=[1, 4, 4, 1],
padding='SAME',
name='z')
with tf.variable_scope('deconv2') as scope:
# Using fiexed bilinearing upsampling filter
w_deconv2 = tf.get_variable('weights',
trainable=False,
initializer=bilinear_upsample_weights(
4, 64))
z_deconv2 = tf.nn.conv2d_transpose(
z_conv2,
w_deconv2, [self.batch_num, shape1, shape2, 64],
strides=[1, 2, 2, 1],
padding='SAME',
name='z')
# Old version API
# merge = tf.concat(3, [conv1_2, z_deconv2, z_deconv3, z_deconv4, z_deconv5])
merge = conv1_2 + z_deconv2 + z_deconv3 + z_deconv4 + z_deconv5
# Add to store dicts
self.outputs['align5'] = z_conv5
self.outputs['align4'] = z_conv4
self.outputs['align3'] = z_conv3
self.outputs['align2'] = z_conv2
self.outputs['deconv5'] = z_deconv5
self.outputs['deconv4'] = z_deconv4
self.outputs['deconv3'] = z_deconv3
self.outputs['deconv2'] = z_deconv2
self.outputs['merge'] = merge
self.layers['align5'] = {'weights': w_conv5, 'biases': b_conv5}
self.layers['align4'] = {'weights': w_conv4, 'biases': b_conv4}
self.layers['align3'] = {'weights': w_conv3, 'biases': b_conv3}
self.layers['align2'] = {'weights': w_conv2, 'biases': b_conv2}
