[dibalokechanda]

Summary

Adds BEPH as a TRIDENT patch encoder.

BEPH is a pathology-pretrained BEiT-v2 Base model. This integration loads the official BEPH backbone through timm by converting the released MMSelfSup/OpenMMLab checkpoint format to the equivalent timm parameter structure.

Changes

• Adds BEPHInferenceEncoder
• Registers --patch_encoder beph
• Adds BEPH to local_ckpts.json
• Converts the official BEPH checkpoint to the timm BEiT parameter layout at load time
• Handles the shared relative-position bias used by the released checkpoint
• Uses ImageNet-normalized 224×224 preprocessing
• Returns 768-dimensional patch embeddings
• Adds an integration test that verifies both the forward pass and embedding dimensionality
• Verified that the preprocessing transform is picklable for DataLoader multiprocessing

Checkpoint download

Unlike several Hugging Face-hosted encoders, BEPH's official pretrained backbone is distributed separately by the authors. Users therefore need to download BEPH_backbone.pth from the official BEPH repository and provide its path either through weights_path or TRIDENT's local_ckpts.json.

For example:

 "beph": "/path/to/BEPH_backbone.pth"

The repository keeps this entry empty by default:

"beph": ""

so no machine-specific checkpoint path is committed.

Testing

TRIDENT_RUN_INTEGRATION_TESTS=1 python -m pytest \
 tests/test_patch_encoders.py -k beph -v

Result:

1 passed

Verified output shape:

 torch.Size([1, 768])

Reference

BEPH: A foundation model for generalizable cancer diagnosis and survival prediction from histopathological images, Nature Communications, 2025.

Official implementation/checkpoint: Zhcyoung/BEPH

[dibalokechanda]

Hi @guillaumejaume just a soft ping on this PR:

It looks like the CI workflow expired while waiting for maintainer approval, so the tests were never able to run. Whenever you have a chance, would you mind approving/retriggering the workflow?

The same failure happened for the other Patch Encoder PRs I made.