deepml.model_arch package

Submodules

deepml.model_arch.dlinknet module

@Credit https://github.com/zlckanata/DeepGlobe-Road-Extraction-Challenge/blob/master/networks/dinknet.py

class deepml.model_arch.dlinknet.Dblock_more_dilate(*args: Any, **kwargs: Any)

Bases: Module

__init__(channel)

forward(x)

class deepml.model_arch.dlinknet.Dblock(*args: Any, **kwargs: Any)

Bases: Module

__init__(channel)

forward(x)

class deepml.model_arch.dlinknet.DecoderBlock(*args: Any, **kwargs: Any)

Bases: Module

__init__(in_channels, n_filters)

forward(x)

class deepml.model_arch.dlinknet.DinkNet34_less_pool(*args: Any, **kwargs: Any)

Bases: Module

__init__(num_classes=1)

forward(x)

class deepml.model_arch.dlinknet.DinkNet34(*args: Any, **kwargs: Any)

Bases: Module

__init__(num_classes=1)

forward(x)

class deepml.model_arch.dlinknet.DinkNet50(*args: Any, **kwargs: Any)

Bases: Module

__init__(num_classes=1)

forward(x)

class deepml.model_arch.dlinknet.DinkNet101(*args: Any, **kwargs: Any)

Bases: Module

__init__(num_classes=1)

forward(x)

class deepml.model_arch.dlinknet.LinkNet34(*args: Any, **kwargs: Any)

Bases: Module

__init__(num_classes=1)

forward(x)

deepml.model_arch.refine_net module

deepml.model_arch.refine_net.convolution_3x3(in_planes, out_planes, stride=1, padding=1, bias=True)

3x3 convolution with padding

deepml.model_arch.refine_net.convolution_1x1(in_planes, out_planes, stride=1, padding=0, bias=True)

1x1 convolution

class deepml.model_arch.refine_net.SpatialAttentionFusionModule(*args: Any, **kwargs: Any)

Bases: Module

Inspired by the attention mechanism, a spatial attention fusion module was designed to enhance useful low-level feature information and remove noise to avoid over using low-level features

__init__()

forward(low_level_features, high_level_features)

Parameters:

• low_level_features – features extracted from backbone

• high_level_features – up sampled features

Returns:

class deepml.model_arch.refine_net.ResidualConvolutionUnit(*args: Any, **kwargs: Any)

Bases: Module

Section 3.2:

The first part of each RefineNet block consists of an adaptive convolution set that mainly fine tunes the pre trained the ResNet weights

__init__(in_planes, out_planes)

forward(x)

class deepml.model_arch.refine_net.MultiResolutionFusion(*args: Any, **kwargs: Any)

Bases: Module

__init__(in_planes, out_planes, fusion_module)

forward(backbone_features, refine_block_features=None)

class deepml.model_arch.refine_net.ChainedResidualPooling(*args: Any, **kwargs: Any)

Bases: Module

Section-1:

Chained residual pooling is able to capture background context from a large image region

__init__(in_planes, out_planes)

forward(x)

class deepml.model_arch.refine_net.RefineBlock(*args: Any, **kwargs: Any)

Bases: Module

__init__(in_planes, out_planes, fusion_module)

forward(backbone_features, refine_block_features=None)

Parameters:

• backbone_features – input from backbone network

• refine_block_features – input from refine net block

Returns:

class deepml.model_arch.refine_net.ReFineNet(*args: Any, **kwargs: Any)

Bases: Module

__init__(res_net_to_use, pre_trained_image_net, top_layers_trainable=True, num_classes=1, fusion_module=False)

refine_block_1

Section 3.2: The final step of Each RefineNet block is another residual convolution unit . This results in a sequence of three RCU between each block. To reflect this behaviour in the last RefineNet-1 block, we place two additional RCU

forward(input_feature)

deepml.model_arch.unet module

deepml.model_arch.unet.convrelu(in_channels, out_channels, kernel, padding)

class deepml.model_arch.unet.ResNetUNet(*args: Any, **kwargs: Any)

Bases: Module

__init__(n_class)

forward(input)

Module contents