search

MhLiao / DB

A PyTorch implementation of "Real-time Scene Text Detection with Differentiable Binarization".
2.08k stars 477 forks source link

Shrink map loss stucked #350

Closed light42 closed 6 months ago

light42 commented 1 year ago

I've tried to train resnet50_db++ by paddleocr. The binary loss(~0.1) and threshold loss(~0.3) decreased significantly, yet only shrink map loss(~0.7) which were stuck at high loss compared to others.

I've tried adjusting learning_rate, weight_decay, unclip_ratio, thresh, box_tresh, shrink_ratio, shrink_thresholds, without any noticeable improvements. In some cases it just make the losses bigger.

Half of my training datasets have width or height that went between 2000-4000 pixels, while the length of bbox could reached 100 pixels.

trugoj commented 8 months ago

Hi there -- any progress on your issue? Did you solve it in the end?

light42 commented 6 months ago

Hi there -- any progress on your issue? Did you solve it in the end?

Sorry, I almost forgot. Its beyond my ability to debug the model so I stopped working on it. My conclusion is both models and my datasets are not adequate enough.

Closing it since it's not my concern anymore.