Skip to content

Missing modules in development/Xenium_classification: custom_dataloader.py, soft_ce.py, convnext.py #5

Description

@RosyGH

Hi STimage team, thank you for open-sourcing this great work — the Nature Communications paper and the code have been very helpful.
I am trying to reproduce the Xenium cell type classification pipeline (the modified HoVer-Net described in the paper, Fig. 4) following the instructions in development/Xenium_classification/README.md. However, the training/inference scripts import three local modules that are not present in the repository:

custom_dataloader.py — imported by both train_hovernet.py and test_hovernet.py (from custom_dataloader import XDataModule)
soft_ce.py — imported by train_hovernet.py (from soft_ce import loss_hovernet_ce, load_df_targets)
convnext.py — imported by both scripts (from convnext import Convnext_HoverNetDecoder)

As a result, both scripts fail immediately with an ImportError, and the README's step 5 (Train model) cannot be executed.
Environment: I set up the conda environment from environment.yml (pathml 2.1.1, PyTorch 2.0.1, CUDA 11.8) on a Linux machine, and successfully ran steps 1–4 (annotation loading, registration, generate_masks.py, preprocess_images.py) — the missing files are the only blocker so far.
Request: Could you please upload these three files (or point me to where they live, if I missed them)? If available, the pretrained HoVer-Net checkpoint used in the paper would also be greatly appreciated, though the missing modules are the critical part.
Happy to help test the pipeline once the files are available. Thanks again for your time!

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions