Recombinator#
- hypercoil.nn.recombinator.Recombinator(in_channels: int, out_channels: int, bias: Tensor = True, positive_only: bool = False, *, key: jax.random.PRNGKey)[source]#
Linear recombinator layer for feature maps. It should also be possible to substitute a 1x1 convolutional layer with similar results.
- Parameters:
- in_channels: int
Number of channels or feature maps input to the recombinator layer.
- out_channels: int
Number of recombined channels or feature maps output by the recombinator layer.
- bias: bool
If True, adds a learnable bias to the output.
- positive_only: bool (default False)
If True, initialise with only positive weights.
- init: dict
Dictionary of parameters to pass to the Kaiming initialisation function. Default:
{'nonlinearity': 'linear'}
- Attributes:
- weight: Tensor
The learnable mixture matrix of the module of shape :math:
C_{in} \times C_{out}.- bias: Tensor
The learnable bias of the module of shape
out_channels.