How to train a multi-task model based on your code?

[o0t1ng0o]

Hi, @haotianliu001 May I ask how to train a multi-task model based on your code?

[haotianll]

Hi, does "multi-task" mean multiple segmentation tasks? If so, I think it's fessible, but you need to modify the code.

This code follows the stardard structure of MMSegmentation. I think you need to determain a suitable label encoding format and write code to load it. It's also necessary to rewrite the loss computation procedure based on this, the testing pipeline here, and the evaluation function here.

[o0t1ng0o]

Thank you for your prompt reply.

My multi-task model may include a segmentation task and a classification task. So I may need to modify the dataset part and add new modules for this. It seems that it is a little bit complicated to modify on your code that is based on mmcv. Could you recommend a multi-lesion segmentation repository that does not based on mmcv?

[haotianll]

It's true. The mmcv framework is a bit complicated to use.

Coincidencely, for the multi-lesion segmentation task, I also have a simpler implementation here, which is based on HRNet-Semantic-Segmentation. This repository is unofficial, but you may try it. And I think it's not hard to add a new classification branch in it.

Besides, you can also find many popular segmentation networks in segmentation_models.pytorch.

[o0t1ng0o]

Thank you very much!!! Your repository is very useful! May I ask can it reach the on-par performance of HRNet that is reported in your paper?

[haotianll]

You are welcome! Yes. They can almost achieve the equal performance.

[o0t1ng0o]

That is goooooood! 😄

[haotianll]

👌