- class flax.linen.Conv(features, kernel_size, strides=1, padding='SAME', input_dilation=1, kernel_dilation=1, feature_group_count=1, use_bias=True, mask=None, dtype=None, param_dtype=<class 'jax.numpy.float32'>, precision=None, kernel_init=<function variance_scaling.<locals>.init>, bias_init=<function zeros>, parent=<flax.linen.module._Sentinel object>, name=None)#
Convolution Module wrapping lax.conv_general_dilated.
Applies a (potentially unshared) convolution to the inputs.
inputs – input data with dimensions (*batch_dims, spatial_dims…, features). This is the channels-last convention, i.e. NHWC for a 2d convolution and NDHWC for a 3D convolution. Note: this is different from the input convention used by lax.conv_general_dilated, which puts the spatial dimensions last. Note: If the input has more than 1 batch dimension, all batch dimensions are flattened into a single dimension for the convolution and restored before returning. In some cases directly vmap’ing the layer may yield better performance than this default flattening approach. If the input lacks a batch dimension it will be added for the convolution and removed n return, an allowance made to enable writing single-example code.
The convolved data.