neuraltrain.models.common.TemporalDownsampling¶

class neuraltrain.models.common.TemporalDownsampling(*, kernel_size: int = 45, stride: int = 45, layer_norm: bool = True, layer_norm_affine: bool = True, gelu: bool = True)[source][source]¶

Temporal downsampling via a 2-D convolution over the time axis.

Parameters:

kernel_size (int) – Kernel height (time-axis) of the downsampling convolution.
stride (int) – Stride along the time axis.
layer_norm (bool) – Apply LayerNorm after the convolution.
layer_norm_affine (bool) – Use learnable affine parameters in the LayerNorm.
gelu (bool) – Apply GELU activation after normalization.