Bug: RDesign's sample function does not actually perform any sampling

[chaitjo]

RDesign/model/rdesign_model.py

Line 76 in fa410a9

def sample(self, X, S, mask=None):

The RDesign model class seems to implement a method called sample(), which is used during inference/testing. However, the sample() method does not actually seem to use any sampling and directly returns raw logits as output.

This is an issue because:

1. This means that any design produced by RDesign is deterministic (one input --> one output sequence) and not diverse.
2. All recovery metric results reported in the paper are not correct (at least based on how recovery is computed in protein design).

Could you clarify whether my understanding is correct, and if so, how one should implementing sampling in RDesign?

For reference, here is how other inverse folding tools implement sampling:

• ProteinMPNN: https://github.com/dauparas/ProteinMPNN/blob/main/protein_mpnn_utils.py#L1180
• ESM-IF: https://github.com/facebookresearch/esm/blob/main/esm/inverse_folding/gvp_transformer.py#L136
• gRNAde: https://github.com/chaitjo/geometric-rna-design/blob/main/src/models.py#L366