def iniatial_block(inputs, name_scope='iniatial_block'):
'''
The initial block for Enet has 2 branches: The convolution branch and Maxpool branch.
The conv branch has 13 filters, while the maxpool branch gives 3 channels corresponding to the RGB channels.
Both output layers are then concatenated to give an output of 16 channels.
:param inputs(Tensor): A 4D tensor of shape [batch_size, height, width, channels]
:return net_concatenated(Tensor): a 4D Tensor of new shape [batch_size, height, width, channels]
'''
# Convolutional branch
with scope(name_scope):
net_conv = conv(inputs, 13, 3, stride=2, padding=1)
net_conv = bn(net_conv)
net_conv = fluid.layers.prelu(net_conv, 'channel')
# Max pool branch
net_pool = max_pool(inputs, [2, 2], stride=2, padding='SAME')
# Concatenated output - does it matter max pool comes first or conv comes first? probably not.
net_concatenated = fluid.layers.concat([net_conv, net_pool], axis=1)
return net_concatenated
def bottleneck(inputs,
output_depth,
filter_size,
regularizer_prob,
projection_ratio=4,
type=REGULAR,
seed=0,
output_shape=None,
dilation_rate=None,
decoder=False,
name_scope='bottleneck'):
# Calculate the depth reduction based on the projection ratio used in 1x1 convolution.
reduced_depth = int(inputs.shape[1] / projection_ratio)
# DOWNSAMPLING BOTTLENECK
if type == DOWNSAMPLING:
#=============MAIN BRANCH=============
#Just perform a max pooling
with scope('down_sample'):
inputs_shape = inputs.shape
with scope('main_max_pool'):
net_main = fluid.layers.conv2d(inputs,
inputs_shape[1],
filter_size=3,
stride=2,
padding='SAME')
#First get the difference in depth to pad, then pad with zeros only on the last dimension.
depth_to_pad = abs(inputs_shape[1] - output_depth)
paddings = [0, 0, 0, depth_to_pad, 0, 0, 0, 0]
with scope('main_padding'):
net_main = fluid.layers.pad(net_main, paddings=paddings)
with scope('block1'):
net = conv(inputs,
reduced_depth, [2, 2],
stride=2,
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
with scope('block2'):
net = conv(net,
reduced_depth, [filter_size, filter_size],
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
with scope('block3'):
net = conv(net, output_depth, [1, 1], padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
# Regularizer
net = fluid.layers.dropout(net, regularizer_prob, seed=seed)
# Finally, combine the two branches together via an element-wise addition
net = fluid.layers.elementwise_add(net, net_main)
net = prelu(net, decoder=decoder)
return net, inputs_shape
# DILATION CONVOLUTION BOTTLENECK
# Everything is the same as a regular bottleneck except for the dilation rate argument
elif type == DILATED:
#Check if dilation rate is given
if not dilation_rate:
raise ValueError('Dilation rate is not given.')
with scope('dilated'):
# Save the main branch for addition later
net_main = inputs
# First projection with 1x1 kernel (dimensionality reduction)
with scope('block1'):
net = conv(inputs, reduced_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
# Second conv block --- apply dilated convolution here
with scope('block2'):
net = conv(net,
reduced_depth,
filter_size,
padding='SAME',
dilation=dilation_rate)
net = bn(net)
net = prelu(net, decoder=decoder)
# Final projection with 1x1 kernel (Expansion)
with scope('block3'):
net = conv(net, output_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
# Regularizer
net = fluid.layers.dropout(net, regularizer_prob, seed=seed)
net = prelu(net, decoder=decoder)
# Add the main branch
net = fluid.layers.elementwise_add(net_main, net)
net = prelu(net, decoder=decoder)
return net
# ASYMMETRIC CONVOLUTION BOTTLENECK
# Everything is the same as a regular bottleneck except for a [5,5] kernel decomposed into two [5,1] then [1,5]
elif type == ASYMMETRIC:
# Save the main branch for addition later
with scope('asymmetric'):
net_main = inputs
# First projection with 1x1 kernel (dimensionality reduction)
with scope('block1'):
net = conv(inputs, reduced_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
# Second conv block --- apply asymmetric conv here
with scope('block2'):
with scope('asymmetric_conv2a'):
net = conv(net,
reduced_depth, [filter_size, 1],
padding='same')
with scope('asymmetric_conv2b'):
net = conv(net,
reduced_depth, [1, filter_size],
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
# Final projection with 1x1 kernel
with scope('block3'):
net = conv(net, output_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
# Regularizer
net = fluid.layers.dropout(net, regularizer_prob, seed=seed)
net = prelu(net, decoder=decoder)
# Add the main branch
net = fluid.layers.elementwise_add(net_main, net)
net = prelu(net, decoder=decoder)
return net
# UPSAMPLING BOTTLENECK
# Everything is the same as a regular one, except convolution becomes transposed.
elif type == UPSAMPLING:
#Check if pooling indices is given
#Check output_shape given or not
if output_shape is None:
raise ValueError('Output depth is not given')
#=======MAIN BRANCH=======
#Main branch to upsample. output shape must match with the shape of the layer that was pooled initially, in order
#for the pooling indices to work correctly. However, the initial pooled layer was padded, so need to reduce dimension
#before unpooling. In the paper, padding is replaced with convolution for this purpose of reducing the depth!
with scope('upsampling'):
with scope('unpool'):
net_unpool = conv(inputs, output_depth, [1, 1])
net_unpool = bn(net_unpool)
net_unpool = fluid.layers.resize_bilinear(
net_unpool, out_shape=output_shape[2:])
# First 1x1 projection to reduce depth
with scope('block1'):
net = conv(inputs, reduced_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
with scope('block2'):
net = deconv(net,
reduced_depth,
filter_size=filter_size,
stride=2,
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
# Final projection with 1x1 kernel
with scope('block3'):
net = conv(net, output_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
# Regularizer
net = fluid.layers.dropout(net, regularizer_prob, seed=seed)
net = prelu(net, decoder=decoder)
# Finally, add the unpooling layer and the sub branch together
net = fluid.layers.elementwise_add(net, net_unpool)
net = prelu(net, decoder=decoder)
return net
# REGULAR BOTTLENECK
else:
with scope('regular'):
net_main = inputs
# First projection with 1x1 kernel
with scope('block1'):
net = conv(inputs, reduced_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
# Second conv block
with scope('block2'):
net = conv(net,
reduced_depth, [filter_size, filter_size],
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
# Final projection with 1x1 kernel
with scope('block3'):
net = conv(net, output_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
# Regularizer
net = fluid.layers.dropout(net, regularizer_prob, seed=seed)
net = prelu(net, decoder=decoder)
# Add the main branch
net = fluid.layers.elementwise_add(net_main, net)
net = prelu(net, decoder=decoder)
return net