Skip to content

v0.3.4 — bugfix: batched residue_ids

Choose a tag to compare

View all tags
@krisss0mecom krisss0mecom released this 01 May 21:08
· 12 commits to main since this release
v0.3.4
f3e37b3

Patch release. Independent third-party test caught that build_features() in alphadynamics.models only accepted residue_ids of shape (N,) — passing a per-batch-item (B, N) tensor raised RuntimeError.

Fix

  • build_features now accepts both (N,) and (B, N) shapes.
  • Regression test added: tests/test_model.py::test_residue_ids_accepts_both_1d_and_2d_shapes
  • Pretrained weights and CLI behavior unchanged.

Credit: caught by an independent test run on a clean PyPI install of v0.3.3 (no DNS access to GitHub Releases, so the bug was found in pure-Python smoke tests rather than via predict).

Headline (unchanged)

  • Mean canonical Ramachandran JSD 0.196 vs Microsoft Timewarp 0.468 on the 4AA benchmark (2.39x lower, 3/3 peptide wins) at ~123K parameters total.
Assets 4
Loading