Fine-tuning BiomedParse V2

[sgilani2020]

Hi @theodore-zhao ,

I’m trying to fine-tune BiomedParse V2 on HPC. There’s no local finetune.py or training script in the repo; can this model be fine-tuned without AzureML, or is the AzureML environment required?

Best regards

[theodore-zhao]

Yes. You don't need to setup AzureML to run this locally. Actually, https://github.com/microsoft/BiomedParse/blob/v2/configs/finetune_biomedparse.yaml works for local jobs. The installed azureml environment is just used to host a training framework we developed to do finetuning (we released that package under the aml library). That does not mean you need to setup your subcription on AzureML. Hope this help clarify!

[sgilani2020]

Thanks for your response! I was able to successfully finetune BiomedParse on ultrasound data locally using that config.

[sanjayss34]

Thanks for releasing your work. It seems that forward_train() in the BiomedParse3D model raises a not implemented error. Should that be used in finetuning, and if so, would you be able to provide the implementation for it?

[theodore-zhao]

Hi, thanks for the question. The v2 model is s 2.5D model performing slice by slice segmentation for 3D volumes. Therefore, it trains in 2D, and inference in 3D. The BiomedParse3D model is a inference interface that provides end-to-end 3D segmentation for a volume, using the pretrained 2D model. Hope this is clear!

[sanjayss34]

Thanks for your reply. Could you explain how the data should be preprocessed if we have 3D CT volumes and want to finetune on that data? (since we need to produce 2D data for finetuning) If you are able to provide a script for doing this, that would be ideal. Thanks again!

[theodore-zhao]

Please check this readme for data preparation:

https://github.com/microsoft/BiomedParse/blob/v2/assets/readmes/DATASET.md