README seems outdated / dependency conflicts with Python versions

[antonio-qsneto]

Hello, @theodore-zhao

I tried setting up the environment following the instructions in the README v2, but I ran into some issues with package compatibility:

1. azureml-automl-core==1.36.0 requires Python < 3.9.
2. The BiomedParse requirements (e.g., PyYAML 6.0.1, numpy 1.26+) require Python ≥ 3.9 / 3.10.
3. Modern PyTorch builds also work best with Python ≥ 3.10.

Attempting to install all dependencies in a single Python environment leads to version conflicts.

Additionally, installing PyTorch via:

conda install pytorch torchvision torchaudio -c pytorch

in Python 3.10 results in:

ImportError: .../libtorch_cpu.so: undefined symbol: iJIT_NotifyEvent

It seems to me that installing Torch via conda isn't working correctly.

This seems to indicate a mismatch between the installed PyTorch binaries and the environment.

In inference_colab_demo.ipynb to colarb, provided to v1, the same error persists with regard to different Python versions:

ERROR: Could not find a version that satisfies the requirement azureml-acft-image-components==0.0.79 (from versions: 0.0.0b1)
ERROR: No matching distribution found for azureml-acft-image-components==0.0.79

Suggestion: In addition to the containerized image (biomedparse_alldata.tar.gz) provided for direct inference, it would be very helpful if the project offered an official Docker setup for the full environment to v2 and how make the inference to biomedparse_alldata.tar.gz.

If anyone has a fork with this fixed or a version with a Docker file, let me know.

Thanks for your attention!

[theodore-zhao]

Thanks for the suggestion! There is an instruction from the challenge website https://www.codabench.org/competitions/5651/#/pages-tab

docker load -i biomedparse_alldata.tar.gz
docker container run --gpus "device=0" -m 32G --name biomedparse --rm -v $PWD/PathToTestSet/:/workspace/inputs/ -v $PWD/biomedparse_outputs/:/workspace/outputs/ biomedparse:latest /bin/bash -c "sh predict.sh"

If this one works nicely for your case, I can add this to the readme so others can use this as an option. Thanks!