How to train?

[wzg722]

How to train?

[github-actions]

Hi @{{ sender.login }}! 👋

Welcome to DTrOCR project! Thank you for your first issue. We're excited to have you here and will respond at the soonest. Happy coding! 🎉

[DemissieDaniel]

@arvindrajan92 how to train?

[arvindrajan92]

@arvindrajan92 how to train?

Hi @DemissieDaniel

Would this notebook be of use: https://github.com/arvindrajan92/DTrOCR/blob/main/examples/IAMFineTune.ipynb.

You would need to build the dataset using the resources outlined in the original DTrOCR paper.

Hope this helps.

[PanFei748]

@arvindrajan92 how to train?

Hi @DemissieDaniel

Would this notebook be of use: https://github.com/arvindrajan92/DTrOCR/blob/main/examples/IAMFineTune.ipynb.

You would need to build the dataset using the resources outlined in the original DTrOCR paper.

Hope this helps.

Hi，why does it perform well on the training set, but very poorly on the validation set, with no decrease in loss

[arvindrajan92]

@arvindrajan92 how to train?

Hi @DemissieDaniel
Would this notebook be of use: https://github.com/arvindrajan92/DTrOCR/blob/main/examples/IAMFineTune.ipynb.
You would need to build the dataset using the resources outlined in the original DTrOCR paper.
Hope this helps.

Hi，why does it perform well on the training set, but very poorly on the validation set, with no decrease in loss

@PanFei748 that would be because the model is not a pre-trained one (like it is mentioned in the paper). I am still working on it on my spare time therefore the progress is slow. Hope I can get a version of pre-trained weights across soon.

[PanFei748]

@arvindrajan92 how to train?

Hi @DemissieDaniel
Would this notebook be of use: https://github.com/arvindrajan92/DTrOCR/blob/main/examples/IAMFineTune.ipynb.
You would need to build the dataset using the resources outlined in the original DTrOCR paper.
Hope this helps.

Hi，why does it perform well on the training set, but very poorly on the validation set, with no decrease in loss

@PanFei748 that would be because the model is not a pre-trained one (like it is mentioned in the paper). I am still working on it on my spare time therefore the progress is slow. Hope I can get a version of pre-trained weights across soon.

First of all, thank you very much for your work. The loss of train data < 0.04 and the accuracy > 98.9, but in val data, the loss remains around 2.7 and the accuracy is around 0.7, thank you for your work again.

[ducanh841988]

@arvindrajan92 could you double check on the image size? I saw you use very small size of image.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.