Lei Xu

Hmm, so it means we lose pytorch random shuffling again? What should be the action items here?

TODO: Verify that `rikai.torch.Dataset` can work with new PyTorch's shuffling capability (or add back shuffling ourselves?)

https://pytorch.org/blog/pytorch-1.11-released/ Probably need to implement this new API

In general, I'd prefer to write `udf` in scala if possible, for two reasons: 1. raw performance 2. It makes udf easier to be available in none python environment, especially...

Oh, let me clarify a bit, i was not suggesting to implement Scala UDF. I was suggesting to write the function directly in `RikaiSparkExtensions` like https://github.com/eto-ai/rikai/blob/main/src/main/scala/ai/eto/rikai/sql/spark/RikaiSparkSessionExtensions.scala#L41-L51 With this, such functions...

This breaks the normal development procedure. Can we either make an instruction in README, or make the docker run part of `pytest`? @Renkai

This happens when I just run `pytest` on my laptop without https://github.com/eto-ai/rikai/blob/master/.github/workflows/python.yml#L40 The instruction of running `docker-compose` is missing. Or better, if we can do run tests without **_manually_** `docker-compose...

What do you think, @darcy-shen ? is this a reasonable ask for spark?

Can this work? https://github.com/eto-ai/rikai/blob/master/python/rikai/types/geometry.py#L211-L223 So basically, we can do ``` scaled_bbox = original_bbox / (1.5, 2.3) ```

Oh, for that, our practice usually was that we just store bbox in the relative scale [0, 1]. So bbox value does not change no matter how Image resizes.