search

ziqi-jin / finetune-anything

Fine-tune SAM (Segment Anything Model) for computer vision tasks such as semantic segmentation, matting, detection ... in specific scenarios
MIT License
787 stars 56 forks source link

Performance #15

Closed kkokilep closed 1 year ago

kkokilep commented 1 year ago

Hello,

Thanks for this amazing repo! I was wondering how good of an MIOU you got on the PASCAL validation set. I trained for about 50,000 iterations and got an MIOU of about 4.12. What performance should I be expecting?

kkokilep commented 1 year ago

I discovered an issue in my version of the code where the labels being returned by the dataloader were between 0 and 1 instead of categorical values. This caused the loss to bounce between a fixed range of values instead, instead of decreasing. I'm not sure if this is a general problem or how I was setting paths in the config file, but I thought I should mention this.

ziqi-jin commented 1 year ago

Hello,

Thanks for this amazing repo! I was wondering how good of an MIOU you got on the PASCAL validation set. I trained for about 50,000 iterations and got an MIOU of about 4.12. What performance should I be expecting?

Hi, the MIOU is from 0 to 100, so 4.12 may not be a good perfomance

ziqi-jin commented 1 year ago

I discovered an issue in my version of the code where the labels being returned by the dataloader were between 0 and 1 instead of categorical values. This caused the loss to bounce between a fixed range of values instead, instead of decreasing. I'm not sure if this is a general problem or how I was setting paths in the config file, but I thought I should mention this.

the 0/1 problem is happened on predictoin label or ground-truth label？

kkokilep commented 1 year ago

It is happening on ground truth labels.

ziqi-jin commented 1 year ago

It is happening on ground truth labels.

Oh, Thanks for your test, I will check it

ziqi-jin commented 1 year ago

It is happening on ground truth labels.

you are right, thank you, and I have fixed the bugs in the latest version. The reason is the label was applied to a ToTensor transform function.