How to trainning with own dataset that customized from Audioset ?

[phungthengoc]

[SarthakYadav]

Hi @phungthengoc, In the current state of the repository, my experiments on AudioSet failed to converge. I'm working on a major update and rework, which streamlines the data loading pipeline, and initial results finally show convergence on AudioSet. Will release complete instructions to train your own model then.

I would recommend that you wait! PS: LEAF implementation itself has nothing to do be convergence issues on AudioSet, so feel free to use that segment. It's data pipeline/hyper-param related issues, which took time to resolve because I moved to a binarized data format (that you'll find in recent commits as packed_dataset.

I'll let the issue remain open! Will definitely get back to you.

[phungthengoc]

Hi @phungthengoc, In the current state of the repository, my experiments on AudioSet failed to converge. I'm working on a major update and rework, which streamlines the data loading pipeline, and initial results finally show convergence on AudioSet. Will release complete instructions to train your own model then.

I would recommend that you wait! PS: LEAF implementation itself has nothing to do be convergence issues on AudioSet, so feel free to use that segment. It's data pipeline/hyper-param related issues, which took time to resolve because I moved to a binarized data format (that you'll find in recent commits as packed_dataset.

I'll let the issue remain open! Will definitely get back to you.

Hi @SarthakYadav , I trained LEAF with tensorflow that sounds good for audio classification task. Moreover, I also want to try with torch-xla about performance, trainning time.