pub fn relpos_encoding(
residue_indices: &Tensor,
n_bins: usize,
) -> Result<Tensor>Expand description
Relative-residue-index one-hot encoding for every (i,j) pair.
For each pair: d = clip(j - i, -n_bins, n_bins), shifted to [0, 2*n_bins],
then one-hot encoded.
§Arguments
residue_indices—[B, N]integer residue positions (any integer dtype)n_bins— half-width of the window (config: 32)
§Returns
[B, N, N, 2*n_bins+1] float32