MBConvBlock¶

tiatoolbox . models . architecture . efficientunet_tissue_mask_model . MBConvBlock

class MBConvBlock(in_planes, out_planes, expand_ratio, kernel_size, stride, reduction_ratio=4)[source]¶

Mobile Inverted Residual Bottleneck block.

This block implements the MBConv block used in EfficientNet architectures. It consists of expansion, depthwise convolution, squeeze-and-excitation, and projection phases with optional residual connection.

use_residual¶

Whether to use residual connection.

Type:: bool

_expand_conv¶

1x1 convolution for channel expansion.

Type:: nn.Module

_bn0¶

Batch normalization after expansion.

Type:: nn.Module

_swish¶

SiLU activation function.

Type:: SiLU

_depthwise_conv¶

Depthwise convolution layer.

Type:: Conv2dStaticSamePadding

_bn1¶

Batch normalization after depthwise convolution.

Type:: nn.BatchNorm2d

_se_reduce¶

Squeeze-and-excitation reduction layer.

Type:: Conv2dStaticSamePadding

_se_expand¶

Squeeze-and-excitation expansion layer.

Type:: Conv2dStaticSamePadding

_project_conv¶

1x1 convolution for projection to output channels.

Type:: Conv2dStaticSamePadding

_bn2¶

Batch normalization after projection.

Type:: nn.BatchNorm2d

Example

>>> block = MBConvBlock(
...     in_planes=32, out_planes=64, expand_ratio=6,
...     kernel_size=3, stride=2
... )
>>> x = torch.randn(1, 32, 64, 64)
>>> output = block(x)
>>> output.shape
... torch.Size([1, 64, 32, 32])

Initialize MBConvBlock.

Creates a mobile inverted residual bottleneck block with expansion, depthwise convolution, squeeze-and-excitation, and projection.

Parameters:

in_planes (int) – Number of input channels.
out_planes (int) – Number of output channels.
expand_ratio (int) – Expansion ratio for the hidden dimension.
kernel_size (int) – Size of the depthwise convolution kernel.
stride (int) – Stride of the depthwise convolution.
reduction_ratio (int) – Reduction ratio for squeeze-and-excitation. Defaults to 4.