Skip to content

Released expert checkpoints are truncated (torch.load fails on all 5) #1

Description

@Hexco

Hi authors, thanks for releasing MoBE.

I downloaded all five files from the linked Drive folder (https://drive.google.com/drive/folders/1YlS66UlnsRlr9QcUKL-EJ_nBfKixnZSk) and verified the byte counts match what Drive serves (so this is not a partial download on my side). None of them can be loaded:

import torch
torch.load("expert_Xray_0.pt", map_location="cpu")
# RuntimeError: PytorchStreamReader failed reading zip archive: failed finding central
# directory. ... high likelihood that your checkpoint file is corrupted.

All five files (Angiogram, CT , MRI , Ultrasound , Xray) are missing the zip End-Of-Central-Directory record, i.e. they were truncated, presumably during upload.

A related observation: since tools/train_experts.py saves the same set of visual-MLP keys for every modality, all five checkpoints should have (nearly) identical file sizes (~227MB in fp32 for BiomedCLIP ViT-B/16's visual MLPs).

Could you please re-upload complete checkpoints (ideally with SHA-256 sums)?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions