def upsample_conv(input_tensor, previous_intermediate_layer, layer_params,
number) -> tf.Tensor:
"""
Deconvolution (upscaling) layers
:param input_tensor:
:param previous_intermediate_layer:
:param layer_params:
:param number:
:return:
"""
with tf.variable_scope('deconv_{}'.format(number)):
if previous_intermediate_layer.get_shape()[1].value and \
previous_intermediate_layer.get_shape()[2].value:
target_shape = previous_intermediate_layer.get_shape()[1:3]
else:
target_shape = tf.shape(previous_intermediate_layer)[1:3]
upsampled_layer = tf.image.resize_images(
input_tensor,
target_shape,
method=tf.image.ResizeMethod.BILINEAR)
net = tf.concat([upsampled_layer, previous_intermediate_layer], 3)
filter_size, nb_bottlenecks = layer_params
if nb_bottlenecks > 0:
for i in range(nb_bottlenecks):
net = resnet_v1.bottleneck(inputs=net,
depth=filter_size,
depth_bottleneck=filter_size //
4,
stride=1)
else:
net = layers.conv2d(inputs=net,
num_outputs=filter_size,
kernel_size=[3, 3],
scope="conv{}".format(number))
return net
def _bottleneck(features, scope):
return resnet_v1.bottleneck(features,
features.shape[-1].value,
features.shape[-1].value,
1,
scope=scope)