This project expired as of February. (Maybe restart after the second half.)
So, I would try to do an experiment that manage this project with yaml and argparser rather than Hydra.
Why
While I use Hydra template, I felt some inconvenience things such as experiment management, directory logging, etc.
The fact of supporting multi-run process is one of the advantage of Hydra template, however, I felt it has some limitations while we add new methods or do experiments and it can be replaced bash file.
Some famous toolkits such as YOLOv8 developed by Ultralytics, timm developed by HuggingFace, don't manage their projects by Hydra template just argparser or yaml. Also, I felt they think the developments' requirements to templates are enough to support DDP(Multi GPU Programming) and not complex directory structure, convenience experiments management.
Especially, management with yaml is a similar way with the mechanism of our benchmark template's inference.
How
It's just an experiment to replace Hydra to yaml and argparser. I'll plan to use yaml main, some default settings would be managed by argparser. I hope and think this paused period makes our template more powerful and robust.
I'll reference these repositories first.
What
This project expired as of February. (Maybe restart after the second half.) So, I would try to do an experiment that manage this project with yaml and argparser rather than Hydra.
Why
While I use Hydra template, I felt some inconvenience things such as experiment management, directory logging, etc. The fact of supporting multi-run process is one of the advantage of Hydra template, however, I felt it has some limitations while we add new methods or do experiments and it can be replaced bash file.
Some famous toolkits such as YOLOv8 developed by Ultralytics, timm developed by HuggingFace, don't manage their projects by Hydra template just argparser or yaml. Also, I felt they think the developments' requirements to templates are enough to support DDP(Multi GPU Programming) and not complex directory structure, convenience experiments management.
Especially, management with yaml is a similar way with the mechanism of our benchmark template's inference.
How
It's just an experiment to replace Hydra to yaml and argparser. I'll plan to use yaml main, some default settings would be managed by argparser. I hope and think this paused period makes our template more powerful and robust. I'll reference these repositories first.