search

open-mmlab / mmocr

OpenMMLab Text Detection, Recognition and Understanding Toolbox
https://mmocr.readthedocs.io/en/dev-1.x/
Apache License 2.0
4.22k stars 739 forks source link

The precision of SDMG-R #430

Closed liangxiaoyun closed 2 years ago

liangxiaoyun commented 2 years ago

Hello, I directly used the config and data(wildreceipt) you provided to train the SDMG-R model and found that it was 4 points lower than your result. Do you know the possible reason?

gaotongxiao commented 2 years ago

We happened to perform a reproducibility test not long ago where we retrained all the models with provided configs (for the release of v0.2.1), and SDMG-R could achieve at least 0.87 f1 score in 4 trials. Have you compared your training log with ours?

liangxiaoyun commented 2 years ago

I used 4 GPUs before, and training with 1 GPU is the same as your results.

amitbcp commented 2 years ago

@liangxiaoyun may be we can scale the batch size when we increase the number of GPU. that should hopefully create a similar result while training on multiple opus

gaotongxiao commented 2 years ago

@liangxiaoyun may be we can scale the batch size when we increase the number of GPU. that should hopefully create a similar result while training on multiple opus

... and tune the learning rate. I guess I can close this issue then.