PyTorch

This page contains the list of all attention models and non-local layers for computer vision enabled with PyTorch backend available in Echo.

Triplet Attention

echoAI.Attention.cv.t_attn.TripletAttention(no_spatial = False, kernel_size = 7)

Triplet Attention

Parameters:

• no_spatial - switches on the spatial attention branch in Triplet Attention. Default: False

• kernel_size - window size of the convolution filters in Triplet Attention. Default: 7

Shape:

• Input:$(\mathbf{N}, \mathbf{C}, \mathbf{H}, \mathbf{W})$4 dimensional feature map tensor.

• Output:$(\mathbf{N}, \mathbf{C}, \mathbf{H}, \mathbf{W})$,same shape as input

Reference:

Rotate to Attend: Convolutional Triplet Attention Module

Squeeze Excite Attention

echoAI.Attention.cv.t_attn.SE(gate_channels, reduction_ratio = 16)

Squeeze Excite Attention

Parameters:

• gate_channels - number of channels in the input tensor. Datatype: Integer

• reduction_ratio - squeeze bottleneck factor of the MLP in Squeeze Excite Attention. Default: 16

Shape:

• Input:$(\mathbf{N}, \mathbf{C}, \mathbf{H}, \mathbf{W})$4 dimensional feature map tensor.

• Output:$(\mathbf{N}, \mathbf{C}, \mathbf{H}, \mathbf{W})$,same shape as input

Reference:

Squeeze-and-Excitation Networks

Convolutional Block Attention Module

echoAI.Attention.cv.t_attn.CBAM(gate_channels, kernel_size = 3, reduction_ratio = 16, pool_types = ['avg', 'max'], no_spatial = False, bam = False, num_layers = 1, bn = False, dilation_conv_num = 2, dilation_val = 4)

Supports both Convolutional Block Attention Module (CBAM) and Bottleneck Attention Module (CBAM)

Convolutional Block Attention Module and Bottleneck Attention Module

Parameters:

• gate_channels - number of channels in the input tensor. Datatype: Integer

• kernel_size - window size of the convolution filters in CBAM/ BAM. Default: 3

• reduction_ratio - width factor of the MLP in CBAM/BAM. Default: 16

• pool_types - listof global pooling operators for channel attention gate in CBAM/BAM. Default: ['avg', 'max']. Note: This is the default for CBAM, which expects two operators, however, if BAM is switched on, pass ['avg']. Available options: avg, lp, max

• no_spatial - switches off the spatial attention gate in CBAM. Default: False

• bam - initializes BAM. Default: False

• num_layers - controls the number of hidden layers in the MLP of channel attention gate in CBAM/BAM. Default: 1

• bn - adds a Batch Normalization layer in the MLP of the channel attention gate in CBAM/BAM. Default: False. Pass True when bam is True.

• dilation_conv_num - number of dilated channel preserving convolution layers in the spatial attention gate in BAM. Default: 2

• dilation_val - dilation factor for the convolution layers in the spatial attention gate in BAM. Default: 4

Note: By default, CBAM is initialized.

Shape:

• Input:$(\mathbf{N}, \mathbf{C}, \mathbf{H}, \mathbf{W})$4 dimensional feature map tensor.

• Output:$(\mathbf{N}, \mathbf{C}, \mathbf{H}, \mathbf{W})$,same shape as input

References:

CBAM: Convolutional Block Attention Module

BAM: Bottleneck Attention Module