table_structure_recognition icon indicating copy to clipboard operation
table_structure_recognition copied to clipboard

train.py file doesn’t exist

Open ssatz opened this issue 1 year ago • 5 comments

Hi, thanks for the repo. I am trying to train the pub1m table structure on yolo8. I couldn’t find the train.py file. Label table row accuracy is no at all good. Any suggestions

ssatz avatar Apr 27 '24 02:04 ssatz

I used Yolov5 to train the model, but I think Yolov8 is similar

whn09 avatar May 27 '24 09:05 whn09

hi thanks this has been solved. I looked into wrong Yolo version. I think you didn't train the Fintab Dataset?

ssatz avatar May 27 '24 11:05 ssatz

Yes, I didn't train the Fintab Dataset. But maybe you can merge the dataset with PubTables-1M. I also recommend you to set Yolo size to at least 800 to get best performance.

whn09 avatar May 28 '24 14:05 whn09

Yes I just trained for fintab with 640 size and table detection is not good and getting short rows. I will try it with 800.

ssatz avatar May 28 '24 14:05 ssatz

I trained both Fintab and PubTables- 1m and normal structure the results are good and i think need to train PaddleOCR for text detection.

Here is the Yolo format data https://huggingface.co/datasets/Codeplug/pub-fintab-yolov Metrics:

image

image

For training the dataset we have used Runpod

ssatz avatar Jun 25 '24 05:06 ssatz