License?

[agoldstein03]

May have missed this, but were the weights/code intended to be released under a specific license? Thanks!

[agoldstein03]

I saw Apache listed for the training data in HF, for when that gets added. I'm hoping that may also be the intent for the code/weights

[jbarrow]

The code is Apache, but I need to check if the model weights can be since they're YOLO models. Currently working on training some Apache-licensed DETRs though. I'll update the license and make a note in the README.

[Zubair983]

The code is Apache, but I need to check if the model weights can be since they're YOLO models. Currently working on training some Apache-licensed DETRs though. I'll update the license and make a note in the README.

Hi @jbarrow, first of all, thanks for your hard work on this project!

I just wanted to follow up and ask if there are any updates on the progress of training the Apache-licensed model? Do you already have some thoughts about it? Are other Apache-licensed base models as accurate as YOLO, or maybe did you hit some problems? Do you maybe already have some timeline in mind when would you release the Apache-licensed model (no pressure, just wanted to check :D )?

Thanks a lot!

[jbarrow]

Just released the Apache licensed model!

Weights are available here: https://huggingface.co/jbarrow/FFDetr

And you can use it with --model FFDetr in the latest commonforms. I'll be pushing an update to PyPI shortly.

[Zubair983]

Thank you @jbarrow very much!!! This is great!

Could you please share what were the training parameters for RF-DETR (number of epochs, image size, same dataset as for FFDNet-L(?))? I see there are differences in models accuracy - YOLO-based model more accurate, less false-positives, so I'm curious if this is because of different training params, or is it just because of changing base model?

Thanks a lot, great work!

[jbarrow]

If you're using the model from within commonforms directly, it's probably because I introduced a bug into the harness at the last moment. I just updated the package to 0.2.1, could you upgrade and retry?

In my experience, the DETR is generally better than the YOLO model (with some cases where it's not).

[Zubair983]

Hi @jbarrow , apologies for the delay.

I compared FFDNet-L & FFDetr results both using library and just a model itself, and TBH I was unable to find cases where DETR would be more accurate that YOLO (for my use cases of course :) ). Overall I'm getting worse fields alignment (shifted or overlapping text in the document), more frequent false-positives. Also FFDetr more frequently detected fields on pages where there was no fields at all (just text and logotypes) where FFDNet-L didn't. Alignment for FFDNet-L is close to pixel perfect, and false-positives are less frequent.

I'm curious what could be the reason for this difference? Were training parameters different for these models? I inspected the FFDetr .pth file and I've seen Namespace(epochs=20), does it mean that FFDetr was trained on 20 epochs, where FFDNet-L was trained on 300 epochs, or is this just about the last iteration? Do you think using different base model (other than RF-DETR) could lead to accuracy closer to YOLO-based?

Below I attach some results from running commonforms cli with 0.3 confidence for both FFDetr & FFDNet-L - inputs are high-res PDF files.

Thank you for your support!