Xingjian Shi

@andreas12345 See https://github.com/awslabs/autogluon/pull/1836 . We are supporting PT 1.11 there.

@andreas12345 now the PT 1.11 has been supported via https://github.com/awslabs/autogluon/pull/1836. Feel free to try installing the latest AutoGluon again (clone it and run `full_install.sh`).

The latest installation guide should have resolved the issue. Close it for now.

@zhiqiangdon has recently implemented the AutoMMPredictor that supports flexible fusion of backbones from TIMM/huggingface as well as numerical / categorical columns

@zhiqiangdon we may need to take a look. There seems to be some issues in ddp_spawn in SageMaker Notebook/Studio environments.

One way to resolve it is to always use "0" as the negative class and use "1" as the positive class.

Should we reuse the `.save()` API?

Usually `.clone()` can be implemented via `.save()` and `.load()`.

A simple solution is to set CUDA_VISIBLE_DEVICES in the environmental variable.

I'll vote for Option 2. `predict_proba` should be disabled if it's a regression problem. Raising an exception will make the life easier and avoid bugs.