Losses and metrics improvements (mostly) - Githubissues

mapbox / robosat

Semantic segmentation on aerial and satellite imagery. Extracts features such as: buildings, parking lots, roads, water, clouds

MIT License

2.02k stars 383 forks source link

Losses and metrics improvements (mostly) #118

Closed ocourtin closed 5 years ago

ocourtin commented 5 years ago

Hi,

Quite a few improvements in this PR:

Add mIoU loss
Add Lovasz loss
Add a config parameter to allow to choose loss function among implemented ones
Remove useless losses parameters (size_average and ignore_index)
Add foreground IoU metric
Add MCC metric
Refactor and cleanup metric class
Add resume training option
Add log training in text file, including hyperparameters

As a consequence:

Imply to add loss parameter in model config file
.pth files already generated are not compliant anymore

Nota:

Pixel accuracy metric wasn't that meaningful, i implement it, than remove it.
mIoU and Lovasz both significantly improve semseg (rather than x entropy loss)

Thanks in advance for the code review ^^

daniel-j-h commented 5 years ago

This looks good to go! Thanks again for these amazing improvements! :tada: :bow:

Could you rebase against master then we can merge this branch.

ocourtin commented 5 years ago

Daniel,

Could you give a look on the rebase result. My first with manuals conflicts to deal with...

Note for myself: create dedicated branches rather than keep working in my own master...

daniel-j-h commented 5 years ago

Looks good; flake8 and black are also fine. Going to merge this in.