How can i train the DOCTR model? How can i create the efficient dataset for train this DOCTR model?
How can i train the DOCTR model? How can i create the efficient dataset for train this DOCTR model?
share me your knowledge about this
Hi,
The model is fine-tuned from the french model. The training is done using only synthetic images. I think this will work quite well with all latin languages.
The training material is made using TextRecognitionDataGenerator
Because of some difficulties installing it I have made a clone of the TextRecognitionDataGenerator at:
https://github.com/diversen/TextRecognitionDataGenerator
In the README you can see instructions on how to generate synthetic images which can then be used as training data.
E.g. if you want to train a model you could use 250000 words and make 8 images for each word.
Then you have 2.000.000 images as training material.
There is also a couple of notes about how to fine-tune a docTR model using this data. This process is also described quite well in the docTR repo.