search

Kohulan / DECIMER-Image-Segmentation

Chemical structure detection and segmentation tool for Journal articles.
https://decimer.ai
MIT License
81 stars 30 forks source link

My training loss looks very good, and the verification loss has been increasing from the beginning. #22

Closed rookiexiao123 closed 3 years ago

rookiexiao123 commented 3 years ago

Can you help me see what the problem is？ 微信图片_20210730132949

rookiexiao123 commented 3 years ago

I tried to reduce the learning rate to 0.0001, which is normal. I also want to split the chemical formula, but my annotation is not very fine. Is the author's annotation as fine as that of the mask RCNN balloon?

OBrink commented 3 years ago

How big are your training and validation datasets? I do not think that the learning rate is the problem here. This looks more like a case of overfitting. The model performs well on the training data but performs poorly on the validation data. My guess is that the training dataset is too small. We used a high-quality dataset of manually annotated pages. I can't really make a clear statement without having more information about your data.

OBrink commented 3 years ago

@rookiexiao123 Has the issue been resolved?

rookiexiao123 commented 3 years ago

I think it is ok. I set the step is 1000,but my dataset size is 160.I think the step set too large.now, I set the step = dataset / batchsize. Loss looks much more normal.

OBrink commented 3 years ago

@rookiexiao123 Thank you for sharing!

rookiexiao123 commented 3 years ago

@OBrink thank you very much!