HTR training issue on input data

[AniketGurav]

Hi,
I have been working on training the HTR (Handwritten Text Recognition) model as per the HTR best practices. However, I am facing some issues during the training process.

I tried training the model using custom data with dimensions (Height: 64, Width: 256, Channels: 3), which I created using the preprocessing script provided by you. Unfortunately, the model is not getting trained properly with these crop sizes. It seems to have convergence issues, and the results are not as expected.

Interestingly, when I trained the model using the original IAM dataset, it converged much faster and performed well. The dimensions of the IAM dataset seem to be different from the custom data I used, but it worked effectively during inference.

I chose the specific sizes (64, 256, 3) for the input data because the output data size during WordStylist inference is expected to be (64, 256, 3).

Could you provide more insights into how the HTR model was trained in the HTR best practices guide? Additionally, could you suggest any adjustments or recommendations for training the model with custom data of size (64, 256, 3) to achieve better convergence and results similar to your paper results?

Your guidance and expertise will be highly appreciated.