search

microsoft / table-transformer

Table Transformer (TATR) is a deep learning model for extracting tables from unstructured documents (PDFs and images). This is also the official repository for the PubTables-1M dataset and GriTS evaluation metric.
MIT License
2.01k stars 231 forks source link

structure fine-tuning issue #181

Open tzktok opened 1 month ago

tzktok commented 1 month ago

Hi, I am trying to fine tune the the table structure over on TATR-v1.1-All-msft.pth model...In my dataset i provide an 50 images for finetune...i am getting below training metrics...

(tables-detr) PS E:\table-transformer\src> python main.py --data_type structure --config_file structure_config.json --data_root_dir data_set --model_load_path TATR-v1.1-All-msft.pth --epochs 50  --mode train --model_save_dir fina
l_model --load_weights_only

pubmed: AP50: 0.811, AP75: 0.637, AP: 0.602, AR: 0.713
Total training time:  0:02:20.607232

How to increase the training accuracy more...i am using default structure config..In dataset how many images needed to get more better result..

@bsmock @dnfclas