search

dbolya / yolact

A simple, fully convolutional model for real-time instance segmentation.
MIT License
5k stars 1.32k forks source link

Which hyperparameters to tune? #246

Open TBdt38 opened 4 years ago

TBdt38 commented 4 years ago

Hi, as i'm running smaller dataset than the full coco + doing fine-tuning, it doesn't take so much time to get reasonable results. So i was thinking to maybe try to run like bayesian optimization to tune some hyperparameters of yolact.

Which hyperparameters do you advise me to release, to run the optimization? which ones are the most influent / interesting? thanks

dbolya commented 4 years ago

Hmm interesting idea. Some things that are particularly influential are the anchor box scales and aspect ratios. Maybe also the number of anchor boxes. Apart from that, there's not much that you can really do a sweep over.

TBdt38 commented 4 years ago

ok nice, i keep that in mind, thanks!

TBdt38 commented 4 years ago

do you have the exact parameter names you mentionned? there are plenty of parameters available in config.py file :) thanks

dbolya commented 4 years ago

They are pred_aspect_ratios and pred_scales. Check yolact_base_config for details.

abhigoku10 commented 4 years ago

@TBondat "run like bayesian optimization to tune some hyperparameters of yolact" can you please elaborate on this is there any different way to tune the hyperparams

TBdt38 commented 4 years ago

yes, you can also do basic grid search or random search for instance. https://nanonets.com/blog/hyperparameter-optimization/