is kolia support python3.6?

[Yuruzu]

[rentruewang]

Sorry, not yet. It doesn't even support python3.7 (#7) haha. We'll get there eventually 😊

[Yuruzu]

SO SAD

[Yuruzu]

Sorry, not yet. It doesn't even support python3.7 (#7) haha. We'll get there eventually 😊

hi, i try to use kolia in python3.8, but i miss trouble too. i follow the getting-started.py and use dgl package to train graph data, but i have the same RuntimeError: CUDA out of memory., can u give a more detailed example of dividing the training, validation, test data set? i should add the code (lazy_input, lazy_label) = lazy(input, label, batch=0) before or after dividing the dataset ? thanks a lot!

[rentruewang]

Hi, thanks for the report!

Koila is a gradient accumulator, so what happens is that is automatically tries to split up the batches (in the background) in the case that there is OOM. However, it doesn't know how to handle "some" layers currently, so it has no effect on those layers. This is done silently in the background. Turns out that if a layer is not recognized by the framework, it doesn't prevent OOM.

As of right now, this project is in a re-write (see #18), which should've happened last month. But I just got a job at a company so currently the schedule is very messy. So I have to apologize for not having time to fix many existing bugs, which would be automatically fixed in a more modular design.

Thanks for reading!

[Yuruzu]

It's okay, looking forward to the next version