Skip to content

CXAS 0.1.1

Choose a tag to compare

@ConstantinSeibold ConstantinSeibold released this 14 Nov 00:44
· 16 commits to main since this release
11af22e

Patch release on top of 0.1.0.

  • Requires transformers>=5.9, the version the models were trained with. With transformers 4.x the frontal model's DINOv3 encoder computes slightly differently, so frontal masks drift: against the training stack on 14 IDC radiographs, 9 structures were kept differently and one structure's mask did not overlap at all. Fresh installs of 0.1.0 already received transformers 5.x and were not affected.
  • Docker images cseibold/cxas:0.1.1-cu128 (also latest; GPU Turing to Blackwell, or CPU) and cseibold/cxas:0.1.1-cpu pin transformers 5.9.0.
  • Model weights unchanged: https://huggingface.co/cmseibold/cxas-models (v0.1.0), https://huggingface.co/cmseibold/cxr-landmarks (v1.0.0), MIT.

Install: pip install cxas==0.1.1. See the 0.1.0 release for the full feature list.