"""
BiSeNet for CelebAMask-HQ, implemented in TensorFlow.
Original paper: 'BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation,'
https://arxiv.org/abs/1808.00897.
"""
__all__ = ['BiSeNet', 'bisenet_resnet18_celebamaskhq']
import os
import tensorflow as tf
import tensorflow.keras.layers as nn
from .common import conv1x1, conv1x1_block, conv3x3_block, InterpolationBlock, MultiOutputSequential, get_channel_axis,\
get_im_size, is_channels_first
from .resnet import resnet18
class PyramidPoolingZeroBranch(nn.Layer):
"""
Pyramid pooling zero branch.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
in_size : tuple of 2 int
Spatial size of output image for the upsampling operation.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
in_channels,
out_channels,
in_size,
data_format="channels_last",
**kwargs):
super(PyramidPoolingZeroBranch, self).__init__(**kwargs)
self.in_size = in_size
self.data_format = data_format
self.pool = nn.GlobalAveragePooling2D(
data_format=data_format,
name="pool")
self.conv = conv1x1_block(
in_channels=in_channels,
out_channels=out_channels,
data_format=data_format,
name="conv")
self.up = InterpolationBlock(
scale_factor=None,
interpolation="bilinear",
data_format=data_format,
name="up")
def call(self, x, training=None):
in_size = self.in_size if self.in_size is not None else get_im_size(x, data_format=self.data_format)
x = self.pool(x)
axis = -1 if is_channels_first(self.data_format) else 1
x = tf.expand_dims(tf.expand_dims(x, axis=axis), axis=axis)
x = self.conv(x, training=training)
x = self.up(x, size=in_size)
return x
class AttentionRefinementBlock(nn.Layer):
"""
Attention refinement block.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
in_channels,
out_channels,
data_format="channels_last",
**kwargs):
super(AttentionRefinementBlock, self).__init__(**kwargs)
self.data_format = data_format
self.conv1 = conv3x3_block(
in_channels=in_channels,
out_channels=out_channels,
data_format=data_format,
name="conv1")
self.pool = nn.GlobalAveragePooling2D(
data_format=data_format,
name="pool")
self.conv2 = conv1x1_block(
in_channels=out_channels,
out_channels=out_channels,
activation="sigmoid",
data_format=data_format,
name="conv2")
def call(self, x, training=None):
x = self.conv1(x, training=training)
w = self.pool(x)
axis = -1 if is_channels_first(self.data_format) else 1
w = tf.expand_dims(tf.expand_dims(w, axis=axis), axis=axis)
w = self.conv2(w, training=training)
x = x * w
return x
class PyramidPoolingMainBranch(nn.Layer):
"""
Pyramid pooling main branch.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
scale_factor : float
Multiplier for spatial size.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
in_channels,
out_channels,
scale_factor,
data_format="channels_last",
**kwargs):
super(PyramidPoolingMainBranch, self).__init__(**kwargs)
self.att = AttentionRefinementBlock(
in_channels=in_channels,
out_channels=out_channels,
data_format=data_format,
name="att")
self.up = InterpolationBlock(
scale_factor=scale_factor,
interpolation="bilinear",
data_format=data_format,
name="up")
self.conv = conv3x3_block(
in_channels=out_channels,
out_channels=out_channels,
data_format=data_format,
name="conv")
def call(self, x, y, training=None):
x = self.att(x, training=training)
x = x + y
x = self.up(x)
x = self.conv(x, training=training)
return x
class FeatureFusion(nn.Layer):
"""
Feature fusion block.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
reduction : int, default 4
Squeeze reduction value.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
in_channels,
out_channels,
reduction=4,
data_format="channels_last",
**kwargs):
super(FeatureFusion, self).__init__(**kwargs)
self.data_format = data_format
mid_channels = out_channels // reduction
self.conv_merge = conv1x1_block(
in_channels=in_channels,
out_channels=out_channels,
data_format=data_format,
name="conv_merge")
self.pool = nn.GlobalAveragePooling2D(
data_format=data_format,
name="pool")
self.conv1 = conv1x1(
in_channels=out_channels,
out_channels=mid_channels,
data_format=data_format,
name="conv1")
self.activ = nn.ReLU()
self.conv2 = conv1x1(
in_channels=mid_channels,
out_channels=out_channels,
data_format=data_format,
name="conv2")
self.sigmoid = tf.nn.sigmoid
def call(self, x, y, training=None):
x = tf.concat([x, y], axis=get_channel_axis(self.data_format))
x = self.conv_merge(x, training=training)
w = self.pool(x)
axis = -1 if is_channels_first(self.data_format) else 1
w = tf.expand_dims(tf.expand_dims(w, axis=axis), axis=axis)
w = self.conv1(w)
w = self.activ(w)
w = self.conv2(w)
w = self.sigmoid(w)
x_att = x * w
x = x + x_att
return x
class PyramidPooling(nn.Layer):
"""
Pyramid Pooling module.
Parameters:
----------
x16_in_channels : int
Number of input channels for x16.
x32_in_channels : int
Number of input channels for x32.
y_out_channels : int
Number of output channels for y-outputs.
y32_out_size : tuple of 2 int
Spatial size of the y32 tensor.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
x16_in_channels,
x32_in_channels,
y_out_channels,
y32_out_size,
data_format="channels_last",
**kwargs):
super(PyramidPooling, self).__init__(**kwargs)
z_out_channels = 2 * y_out_channels
self.pool32 = PyramidPoolingZeroBranch(
in_channels=x32_in_channels,
out_channels=y_out_channels,
in_size=y32_out_size,
data_format=data_format,
name="pool32")
self.pool16 = PyramidPoolingMainBranch(
in_channels=x32_in_channels,
out_channels=y_out_channels,
scale_factor=2,
data_format=data_format,
name="pool16")
self.pool8 = PyramidPoolingMainBranch(
in_channels=x16_in_channels,
out_channels=y_out_channels,
scale_factor=2,
data_format=data_format,
name="pool8")
self.fusion = FeatureFusion(
in_channels=z_out_channels,
out_channels=z_out_channels,
data_format=data_format,
name="fusion")
def call(self, x8, x16, x32, training=None):
y32 = self.pool32(x32, training=training)
y16 = self.pool16(x32, y32, training=training)
y8 = self.pool8(x16, y16, training=training)
z8 = self.fusion(x8, y8, training=training)
return z8, y8, y16
class BiSeHead(nn.Layer):
"""
BiSeNet head (final) block.
Parameters:
----------
in_channels : int
Number of input channels.
mid_channels : int
Number of middle channels.
out_channels : int
Number of output channels.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
in_channels,
mid_channels,
out_channels,
data_format="channels_last",
**kwargs):
super(BiSeHead, self).__init__(**kwargs)
self.conv1 = conv3x3_block(
in_channels=in_channels,
out_channels=mid_channels,
data_format=data_format,
name="conv1")
self.conv2 = conv1x1(
in_channels=mid_channels,
out_channels=out_channels,
data_format=data_format,
name="conv2")
def call(self, x, training=None):
x = self.conv1(x, training=training)
x = self.conv2(x)
return x
class BiSeNet(tf.keras.Model):
"""
BiSeNet model from 'BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation,'
https://arxiv.org/abs/1808.00897.
Parameters:
----------
backbone : func -> nn.Sequential
Feature extractor.
aux : bool, default True
Whether to output an auxiliary results.
fixed_size : bool, default True
Whether to expect fixed spatial size of input image.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (640, 480)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
"""
def __init__(self,
backbone,
aux=True,
fixed_size=True,
in_channels=3,
in_size=(640, 480),
classes=19,
data_format="channels_last",
**kwargs):
super(BiSeNet, self).__init__(**kwargs)
assert (in_channels == 3)
self.in_size = in_size
self.classes = classes
self.data_format = data_format
self.aux = aux
self.fixed_size = fixed_size
self.backbone, backbone_out_channels = backbone(
data_format=data_format,
name="backbone")
y_out_channels = backbone_out_channels[0]
z_out_channels = 2 * y_out_channels
y32_out_size = (self.in_size[0] // 32, self.in_size[1] // 32) if fixed_size else None
self.pool = PyramidPooling(
x16_in_channels=backbone_out_channels[1],
x32_in_channels=backbone_out_channels[2],
y_out_channels=y_out_channels,
y32_out_size=y32_out_size,
data_format=data_format,
name="pool")
self.head_z8 = BiSeHead(
in_channels=z_out_channels,
mid_channels=z_out_channels,
out_channels=classes,
data_format=data_format,
name="head_z8")
self.up8 = InterpolationBlock(
scale_factor=(8 if fixed_size else None),
data_format=data_format,
name="up8")
if self.aux:
mid_channels = y_out_channels // 2
self.head_y8 = BiSeHead(
in_channels=y_out_channels,
mid_channels=mid_channels,
out_channels=classes,
data_format=data_format,
name="head_y8")
self.head_y16 = BiSeHead(
in_channels=y_out_channels,
mid_channels=mid_channels,
out_channels=classes,
data_format=data_format,
name="head_y16")
self.up16 = InterpolationBlock(
scale_factor=(16 if fixed_size else None),
data_format=data_format,
name="up16")
def call(self, x, training=None):
assert is_channels_first(self.data_format) or ((x.shape[1] % 32 == 0) and (x.shape[2] % 32 == 0))
assert (not is_channels_first(self.data_format)) or ((x.shape[2] % 32 == 0) and (x.shape[3] % 32 == 0))
x8, x16, x32 = self.backbone(x, training=training)
z8, y8, y16 = self.pool(x8, x16, x32, training=training)
z8 = self.head_z8(z8, training=training)
z8 = self.up8(z8)
if self.aux:
y8 = self.head_y8(y8, training=training)
y16 = self.head_y16(y16, training=training)
y8 = self.up8(y8)
y16 = self.up16(y16)
return z8, y8, y16
else:
return z8
def get_bisenet(model_name=None,
pretrained=False,
root=os.path.join("~", ".tensorflow", "models"),
**kwargs):
"""
Create BiSeNet model with specific parameters.
Parameters:
----------
model_name : str or None, default None
Model name for loading pretrained model.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
"""
net = BiSeNet(
**kwargs)
if pretrained:
if (model_name is None) or (not model_name):
raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.")
from .model_store import get_model_file
in_channels = kwargs["in_channels"] if ("in_channels" in kwargs) else 3
input_shape = (1,) + (in_channels,) + net.in_size if net.data_format == "channels_first" else\
(1,) + net.in_size + (in_channels,)
net.build(input_shape=input_shape)
net.load_weights(
filepath=get_model_file(
model_name=model_name,
local_model_store_dir_path=root))
return net
def bisenet_resnet18_celebamaskhq(pretrained_backbone=False, classes=19, **kwargs):
"""
BiSeNet model on the base of ResNet-18 for face segmentation on CelebAMask-HQ from 'BiSeNet: Bilateral Segmentation
Network for Real-time Semantic Segmentation,' https://arxiv.org/abs/1808.00897.
Parameters:
----------
pretrained_backbone : bool, default False
Whether to load the pretrained weights for feature extractor.
classes : int, default 19
Number of classes.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
"""
def backbone(**bb_kwargs):
features_raw = resnet18(pretrained=pretrained_backbone, **bb_kwargs).features
features_raw._layers.pop()
features = MultiOutputSequential(return_last=False, name="backbone")
features.add(features_raw._layers[0])
for i, stage in enumerate(features_raw._layers[1:]):
if i != 0:
stage.do_output = True
features.add(stage)
out_channels = [128, 256, 512]
return features, out_channels
return get_bisenet(backbone=backbone, classes=classes, model_name="bisenet_resnet18_celebamaskhq", **kwargs)
def _test():
import numpy as np
import tensorflow.keras.backend as K
data_format = "channels_last"
# data_format = "channels_first"
in_size = (640, 480)
aux = True
pretrained = False
models = [
bisenet_resnet18_celebamaskhq,
]
for model in models:
net = model(pretrained=pretrained, in_size=in_size, aux=aux, data_format=data_format)
batch = 14
x = tf.random.normal((batch, 3, in_size[0], in_size[1]) if is_channels_first(data_format) else
(batch, in_size[0], in_size[1], 3))
ys = net(x)
y = ys[0] if aux else ys
assert (y.shape[0] == x.shape[0])
if is_channels_first(data_format):
assert ((y.shape[1] == 19) and (y.shape[2] == x.shape[2]) and (y.shape[3] == x.shape[3]))
else:
assert ((y.shape[3] == 19) and (y.shape[1] == x.shape[1]) and (y.shape[2] == x.shape[2]))
weight_count = sum([np.prod(K.get_value(w).shape) for w in net.trainable_weights])
print("m={}, {}".format(model.__name__, weight_count))
if aux:
assert (model != bisenet_resnet18_celebamaskhq or weight_count == 13300416)
else:
assert (model != bisenet_resnet18_celebamaskhq or weight_count == 13150272)
if __name__ == "__main__":
_test()
