def fcn8sd_resnetd50b_cityscapes(pretrained_backbone=False, classes=19, aux=True, data_format="channels_last",
**kwargs):
"""
FCN-8s(d) model on the base of ResNet(D)-50b for Cityscapes from 'Fully Convolutional Networks for Semantic
Segmentation,' https://arxiv.org/abs/1411.4038.
Parameters:
----------
pretrained_backbone : bool, default False
Whether to load the pretrained weights for feature extractor.
classes : int, default 19
Number of segmentation classes.
aux : bool, default True
Whether to output an auxiliary result.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
"""
backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,),
data_format=data_format).features
backbone.children.pop()
return get_fcn8sd(backbone=backbone, classes=classes, aux=aux, model_name="fcn8sd_resnetd50b_cityscapes",
data_format=data_format, **kwargs)
def pspnet_resnetd50b_ade20k(pretrained_backbone=False, classes=150, aux=True, data_format="channels_last", **kwargs):
"""
PSPNet model on the base of ResNet(D)-50b for ADE20K from 'Pyramid Scene Parsing Network,'
https://arxiv.org/abs/1612.01105.
Parameters:
----------
pretrained_backbone : bool, default False
Whether to load the pretrained weights for feature extractor.
classes : int, default 150
Number of segmentation classes.
aux : bool, default True
Whether to output an auxiliary result.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
"""
backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,),
data_format=data_format).features
backbone.children.pop()
return get_pspnet(backbone=backbone, classes=classes, aux=aux, model_name="pspnet_resnetd50b_ade20k",
data_format=data_format, **kwargs)
def deeplabv3_resnetd50b_ade20k(pretrained_backbone=False, classes=150, aux=True, data_format="channels_last",
**kwargs):
"""
DeepLabv3 model on the base of ResNet(D)-50b for ADE20K from 'Rethinking Atrous Convolution for Semantic Image
Segmentation,' https://arxiv.org/abs/1706.05587.
Parameters:
----------
pretrained_backbone : bool, default False
Whether to load the pretrained weights for feature extractor.
classes : int, default 150
Number of segmentation classes.
aux : bool, default True
Whether to output an auxiliary result.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
"""
backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,),
data_format=data_format).features
backbone.children.pop()
return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd50b_ade20k",
data_format=data_format, **kwargs)
def icnet_resnetd50b_cityscapes(pretrained_backbone=False,
classes=19,
aux=True,
data_format="channels_last",
**kwargs):
"""
ICNet model on the base of ResNet(D)-50b for Cityscapes from 'ICNet for Real-Time Semantic Segmentation on
High-Resolution Images,' https://arxiv.org/abs/1704.08545.
Parameters:
----------
pretrained_backbone : bool, default False
Whether to load the pretrained weights for feature extractor.
classes : int, default 19
Number of segmentation classes.
aux : bool, default True
Whether to output an auxiliary result.
data_format : str, default 'channels_last'
The ordering of the dimensions in tensors.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.tensorflow/models'
Location for keeping the model parameters.
"""
backbone1 = resnetd50b(pretrained=pretrained_backbone,
ordinary_init=False,
bends=None,
data_format=data_format).features
for i in range(len(backbone1) - 3):
backbone1.children.pop()
backbone2 = resnetd50b(pretrained=pretrained_backbone,
ordinary_init=False,
bends=None,
data_format=data_format).features
backbone2.children.pop()
for i in range(3):
backbone2.children.pop(0)
backbones = (backbone1, backbone2)
backbones_out_channels = (512, 2048)
return get_icnet(backbones=backbones,
backbones_out_channels=backbones_out_channels,
classes=classes,
aux=aux,
model_name="icnet_resnetd50b_cityscapes",
data_format=data_format,
**kwargs)
