Squeeze and Excitation

This module is inspired by the Squeeze and Excitation Normalization module: github or arxiv

Class - SqueezeAndExcitation2D

Simple SqueezeAndExcitation for 2D filters.

Parameter	type	Description
in_channels	int	Number of input channels
reduction	int, default=2	Degree of reduction

Return Value

The returned value is the squeeze and excitation map. This map is not applied to the original input in any way.

Example

A simple usage example without context:

import torch
import useful_layers as ul

class demo(torch.nn.Module):
    def __init__(self, in_channels):
        # ...
        self.layer = ul.layers.SqueezeAndExcitation2D(in_channels=in_channels,
                                                      reduction=2)

    def forward(self, x):
        se_map = self.layer(x)   # This map is not applied!
        return x * se_map        # This is the default behaviour as described in the paper

Class - SqueezeAndExcitation3D

Simple SqueezeAndExcitation for 3D filters.

Parameter	type	Description
in_channels	int	Number of input channels
reduction	int, default=2	Degree of reduction

Return Value

The returned value is the Squeeze and Excitation map. This map is not applied to the original input in any way.

Example

A simple usage example without context:

import torch
import useful_layers as ul

class demo(torch.nn.Module):
    def __init__(self, in_channels):
        # ...
        self.layer = ul.layers.SqueezeAndExcitation3D(in_channels=in_channels,
                                                      reduction=2)

    def forward(self, x):
        se_map = self.layer(x)   # This map is not applied!
        return x * se_map        # This is the default behaviour as described in the paper