Result are worse after fine-tune

147 views
Skip to first unread message

Omar Samir

unread,
Mar 3, 2024, 5:43:14 AM3/3/24
to tesseract-ocr
I was fine-tuning the ara.traineddata in tessdata_best and I have benchmarked the output model and tessdata_best model. I have found that the tessdata_best model did better than the output model, so is it a normal thing that my output model did worse? If not what am I doing wrong? 

I am using Tesseract 5.3.3 and ubuntu wsl.

This is the training line:

```make LANG_TYPE=RTL MODEL_NAME=Tesseract_F_JSTOR_TRDG_Shaden PSM=13 START_MODEL=ara TESSDATA=../tessdata_best FINETUNE_TYPE=Layer MAX_ITERATIONS=21000 RATIO_TRAIN=0.80 EPOCHS=3 training -j12``` 
Reply all
Reply to author
Forward
0 new messages