search

Layout-Parser / layout-model-training

The scripts for training Detectron2-based Layout Models on popular layout analysis datasets
198 stars 54 forks source link

Unable to reproduce performance #26

Open y2sman opened 2 months ago

y2sman commented 2 months ago 
,bbox,segm
AP,32.66216488648993,31.699710987990564
AP50,44.651233552341395,44.53845613640047
AP75,36.13727520014585,35.81650765346808
APs,25.210783194288627,22.183014770683627
APm,18.897219529038455,19.1843523461893
APl,37.33546062498339,38.234935520576855
AP-Background,,
AP-TextRegion,65.25687119406678,65.77834127648671
AP-ImageRegion,59.93286761496787,63.302475927257476
AP-TableRegion,37.07920792079208,37.07920792079208
AP-MathsRegion,28.58085808580858,19.05940594059406
AP-SeparatorRegion,2.0847340495051303,1.3397419212608506
AP-OtherRegion,3.0384504537991393,3.6390929415522195

Thanks for your brilliant work.

I tried perform training by following the "How to train the models?" written on README.

I suffered a problem named "FloatingPointError: Predicted boxes or scores contain Inf/NaN. Training has diverged." and it is also reported in #21. So, i changed BASE_LR 0.02 to 0.01.

"bbox AP" should be 69.35377194762384 reported to https://layout-parser.readthedocs.io/en/latest/notes/modelzoo.html , but it is not.

Can you provide the exact version of the libraries for reproduction?