Masked Attention
- class complextorch.nn.modules.attention.mca.CVMaskedChannelAttention1d(channels: int, reduction_factor: int = 2, MaskingClass: ~torch.nn.modules.module.Module = <class 'complextorch.nn.modules.mask.ComplexRatioMask'>, act: ~torch.nn.modules.module.Module = <class 'complextorch.nn.modules.activation.complex_relu.CReLU'>)
1-D Complex-Valued Masked Channel Attention (MCA) Module
Generalized for arbitrary masking function (see mask for implemented masking functions)
Implements the operation:
\[\texttt{CV-MCA}(\mathbf{z}) = \mathcal{M}(H_\text{ConvUp}(\mathcal{A}(H_\text{ConvDown}(\mathbf{z})))) \odot \mathbf{z},\]where \(\mathcal{M}(\cdot)\) is the masking function (by default, ComplexRatioMask is used) and \(H_\text{ConvUp}(\cdot)\) and \(H_\text{ConvDown}(\cdot)\) are 1-D convolution layers with kernel sizes of 1 that reduce the channel dimension by a factor \(r\).
Based on work from the following paper:
HW Cho, S Choi, YR Cho, and J Kim: Complex-Valued Channel Attention and Application in Ego-Velocity Estimation With Automotive Radar
- class complextorch.nn.modules.attention.mca.CVMaskedChannelAttention2d(channels: int, reduction_factor: int = 2, MaskingClass: ~torch.nn.modules.module.Module = <class 'complextorch.nn.modules.mask.ComplexRatioMask'>, act: ~torch.nn.modules.module.Module = <class 'complextorch.nn.modules.activation.complex_relu.CReLU'>)
2-D Complex-Valued Masked Channel Attention (MCA) Module
Implements the operation:
\[\texttt{CV-MCA}(\mathbf{z}) = \mathcal{M}(H_\text{ConvUp}(\mathcal{A}(H_\text{ConvDown}(\mathbf{z})))) \odot \mathbf{z},\]where \(\mathcal{M}(\cdot)\) is the masking function (by default, ComplexRatioMask is used) and \(H_\text{ConvUp}(\cdot)\) and \(H_\text{ConvDown}(\cdot)\) are 2-D convolution layers with kernel sizes of 1 that reduce the channel dimension by a factor \(r\).
Generalized for arbitrary masking function (see mask for implemented masking functions)
Based on work from the following paper:
HW Cho, S Choi, YR Cho, and J Kim: Complex-Valued Channel Attention and Application in Ego-Velocity Estimation With Automotive Radar
- class complextorch.nn.modules.attention.mca.CVMaskedChannelAttention3d(channels: int, reduction_factor: int = 2, MaskingClass: ~torch.nn.modules.module.Module = <class 'complextorch.nn.modules.mask.ComplexRatioMask'>, act: ~torch.nn.modules.module.Module = <class 'complextorch.nn.modules.activation.complex_relu.CReLU'>)
3-D Complex-Valued Masked Channel Attention (MCA) Module
Generalized for arbitrary masking function (see mask for implemented masking functions)
Implements the operation:
\[\texttt{CV-MCA}(\mathbf{z}) = \mathcal{M}(H_\text{ConvUp}(\mathcal{A}(H_\text{ConvDown}(\mathbf{z})))) \odot \mathbf{z},\]where \(\mathcal{M}(\cdot)\) is the masking function (by default, ComplexRatioMask is used) and \(H_\text{ConvUp}(\cdot)\) and \(H_\text{ConvDown}(\cdot)\) are 3-D convolution layers with kernel sizes of 1 that reduce the channel dimension by a factor \(r\).
Based on work from the following paper:
HW Cho, S Choi, YR Cho, and J Kim: Complex-Valued Channel Attention and Application in Ego-Velocity Estimation With Automotive Radar