- neural_tangents.stax.ConvLocal(out_chan, filter_shape, strides=None, padding='VALID', W_std=1.0, b_std=None, dimension_numbers=None, parameterization='ntk', s=(1, 1))[source]
General unshared convolution.
Also known and “Locally connected networks” or LCNs, these are equivalent to convolutions except for having separate (unshared) kernels at different spatial locations.
int) – The number of output channels / features of the convolution. This is ignored in by the
int]) – The shape of the filter. The shape of the tuple should agree with the number of spatial dimensions in
int]]) – The stride of the convolution. The shape of the tuple should agree with the number of spatial dimensions in
str) – Specifies padding for the convolution. Can be one of
"CIRCULAR"uses periodic convolutions.
float) – standard deviation of the weights.
float]) – standard deviation of the biases.
Nonemeans no bias.
str]]) – Specifies which axes should be convolved over. Should match the specification in
str) – Either
"standard". These parameterizations are the direct analogues for convolution of the corresponding parameterizations for
int]) – A tuple of integers, a direct convolutional analogue of the respective parameters for the
- Return type
(init_fn, apply_fn, kernel_fn).