def __init__(self, filters, activation, y_conditioned=False, splits=None):
super(RefineNet, self).__init__()
self.in_shape = None
# Boolean is True when conditional information is concatenated to x
self.y_conditioned = y_conditioned
self.splits = splits
self.increase_channels = layers.Conv2D(filters,
kernel_size=3,
padding="same")
self.preact_1 = ConditionalFullPreActivationBlock(activation,
filters,
kernel_size=3)
# FIXME: In this second preactivation block we used dilation=1, but it should have been 2
self.preact_2 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
pooling=True,
dilation=2,
padding=2)
self.preact_3 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
dilation=2,
padding=2)
self.preact_4 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
dilation=4,
padding=4)
self.refine_block_1 = RefineBlock(
activation,
filters,
n_blocks_crp=2,
n_blocks_begin_rcu=2,
n_blocks_end_rcu=3,
)
self.refine_block_2 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.refine_block_3 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.refine_block_4 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.norm = ConditionalInstanceNormalizationPlusPlus2D()
self.activation = activation
self.decrease_channels = None
def __init__(self, filters, activation):
super(ResNet, self).__init__()
self.increase_channels = layers.Conv2D(filters,
kernel_size=3,
padding='same')
self.res_enc1 = ConditionalResidualBlock(activation,
filters,
is_encoder=True)
self.res_enc2 = ConditionalResidualBlock(activation,
filters * 2,
dilation=2,
is_encoder=True,
resize=True)
self.res_enc3 = ConditionalResidualBlock(activation,
filters * 2,
dilation=2,
is_encoder=True)
self.res_enc4 = ConditionalResidualBlock(activation,
filters * 4,
dilation=2,
is_encoder=True,
resize=True)
self.res_enc5 = ConditionalResidualBlock(activation,
filters * 4,
dilation=2,
is_encoder=True)
self.res_dec1 = ConditionalResidualBlock(activation,
filters * 4,
dilation=2,
is_encoder=False)
self.res_dec2 = ConditionalResidualBlock(activation,
filters * 4,
dilation=2,
is_encoder=False,
resize=True)
self.res_dec3 = ConditionalResidualBlock(activation,
filters * 2,
dilation=2,
is_encoder=False)
self.res_dec4 = ConditionalResidualBlock(activation,
filters * 2,
dilation=2,
is_encoder=False,
resize=True)
self.res_dec5 = ConditionalResidualBlock(activation,
filters,
is_encoder=False)
self.norm = ConditionalInstanceNormalizationPlusPlus2D()
self.activation = activation
self.decrease_channels = None
def __init__(self, filters, activation):
super(RefineNet, self).__init__()
self.in_shape = None
self.increase_channels = layers.Conv2D(filters,
kernel_size=3,
padding='same')
self.preact_1 = ConditionalFullPreActivationBlock(activation,
filters,
kernel_size=3)
# FIXME: In this second preactivation block we used dilation=1, but it should have been 2
self.preact_2 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
pooling=True)
self.preact_3 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
dilation=2,
padding=2)
self.preact_4 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
dilation=4,
padding=4)
self.refine_block_1 = RefineBlock(activation,
filters,
n_blocks_crp=2,
n_blocks_begin_rcu=2,
n_blocks_end_rcu=3)
self.refine_block_2 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.refine_block_3 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.refine_block_4 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.norm = ConditionalInstanceNormalizationPlusPlus2D()
self.activation = activation
self.decrease_channels = None
def __init__(self,
activation,
filters,
is_encoder,
kernel_size=3,
dilation=1,
resize=False):
super(ConditionalResidualBlock, self).__init__()
self.norm1 = ConditionalInstanceNormalizationPlusPlus2D()
self.norm2 = ConditionalInstanceNormalizationPlusPlus2D()
self.activation = activation
self.resize = resize
self.filters = filters
self.is_encoder = is_encoder
if is_encoder:
self.conv1 = layers.Conv2D(filters,
kernel_size,
dilation_rate=(dilation, dilation),
padding="same")
self.conv2 = layers.Conv2D(filters,
kernel_size,
dilation_rate=(dilation, dilation),
padding="same")
else:
self.conv1 = layers.Conv2DTranspose(filters,
kernel_size,
dilation_rate=(dilation,
dilation),
padding="same")
self.conv2 = layers.Conv2DTranspose(filters,
kernel_size,
dilation_rate=(dilation,
dilation),
padding="same")
self.adjust_skip = None
def __init__(self, filters, activation, splits, y_conditioned=False):
super(MaskedRefineNet, self).__init__()
self.in_shape = None
self.filters = filters
# Describes how to split the conditioning information
# Assumes x is concatenated with its pixel-wise conditioning info
self.splits = splits
# Concatenate x,y too feed into RefineNet block
# Encourages local per-pixel conditioning
self.y_conditioned = y_conditioned
# Filters used for every channel when dotting
# dotting with conditioning block
self.conditional_embedding_size = 256
self.increase_channels = layers.Conv2D(filters,
kernel_size=3,
padding="same")
self.preact_1 = ConditionalFullPreActivationBlock(activation,
filters,
kernel_size=3)
# FIXME: In this second preactivation block we used dilation=1, but it should have been 2
self.preact_2 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
pooling=True)
self.preact_3 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
dilation=2,
padding=2)
self.preact_4 = ConditionalFullPreActivationBlock(activation,
filters * 2,
kernel_size=3,
dilation=4,
padding=4)
self.refine_block_1 = RefineBlock(
activation,
filters,
n_blocks_crp=2,
n_blocks_begin_rcu=2,
n_blocks_end_rcu=3,
)
self.refine_block_2 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.refine_block_3 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.refine_block_4 = RefineBlock(activation,
filters * 2,
n_blocks_crp=2,
n_blocks_begin_rcu=2)
self.norm = ConditionalInstanceNormalizationPlusPlus2D()
self.activation = activation
self.decrease_channels = None
self.depth_extension = layers.Conv2D(
self.conditional_embedding_size * self.splits[0],
kernel_size=3,
padding="same",
)
# Added for conditioning block
self.increase_conditional_channels = layers.Conv2D(filters,
kernel_size=3,
padding="same")
# self.conditional_block = FullPreActivationBlock(activation, self.conditional_embedding_size*splits[0], kernel_size=3)
self.conditional_block = FullPreActivationBlock(
activation,
self.conditional_embedding_size,
kernel_size=3,
dilation=2,
padding=2,
)
self.pool = GlobalAveragePooling2D(data_format="channels_last")
self.dot = tf.keras.layers.Dot(axes=(4, 1), name="condition_dot")