open-mmlab / mmsegmentation

OpenMMLab Semantic Segmentation Toolbox and Benchmark.
https://mmsegmentation.readthedocs.io/en/main/
Apache License 2.0
7.88k stars 2.56k forks source link

Validation Metrics Don't Save to WandB Via MMSegWandbHook #2236

Open AndrewTKent opened 1 year ago

AndrewTKent commented 1 year ago

Motivation I am able to see the training loss metrics logged into my WandB account using MMSegWandbHook, but I'd much prefer to see validation metrics like mIoU or mAcc across iterations - even if I am only able to see them during iterations where I am preforming an evaluation.

To my knowledge MMSegWandbHook doesn't allow for this, and I've tried looking -

MeowZheng commented 1 year ago

Might try to modify the evaluation configuration and shorten the interval between evaluations, as wandb will record the evaluation results every evaluation iter. https://github.com/open-mmlab/mmsegmentation/blob/8dbbdd8c17a082c63947e6d20349e751b093f640/configs/_base_/schedules/schedule_40k.py#L9 https://github.com/open-mmlab/mmsegmentation/blob/8dbbdd8c17a082c63947e6d20349e751b093f640/mmseg/core/hook/wandblogger_hook.py#L143

AndrewTKent commented 1 year ago

@MeowZheng I have my evaluation set to 100, and I'm able to see the validation metrics in the log from TextLoggerHook but nothing is showing up in my WandB. Should I paste my config? Thank-you for helping me with this

MeowZheng commented 1 year ago

hi @ayulockin please take a look, thank you.