[models] Add VIPTR recognition model

[felixdittrich92]

🚀 The feature

Ref.: #1826

Paper: VIPTR
Implementation: https://github.com/cxfyxl/VIPTR

• PyTorch implementation
• TensorFlow implementation

Hi, 

I would like to suggest possibly introducing another state-of-the-art text recognition architecture to docTR.
[SVIPTR](https://paperswithcode.com/paper/viptr-a-vision-permutable-extractor-for-fast)
It's promising accurate results at low latency.

Notably, the SVIPTR-T (Tiny) variant delivers highly competitive accuracy on par with other lightweight models and achieves SOTA inference speeds. Meanwhile, the SVIPTR-L (Large) attains SOTA accuracy in single-encoder-type models, while maintaining a low parameter count and favorable inference speed.

Thanks for your consideration.

Inference latency should be comparable to crnn_mobilenet_v3_large and the results are hopefully comparable to parseq.
The addition is agreed.

[felixdittrich92]

If someone wants to work on this feel free to ping here. Otherwise I planned to start working on it after we have some strategy done to make docTR multilingual.

[lkosh]

Hi, if the issue is still open, I'd like to work on the pytorch implementation of this feature. I've been using doctr quite extensively lately and I'd like to help improving this wonderful project :)

[felixdittrich92]

Hi, if the issue is still open, I'd like to work on the pytorch implementation of this feature. I've been using doctr quite extensively lately and I'd like to help improving this wonderful project :)

Hey @lkosh 👋

Sure highly apprecated 👍

I started already with the PT implementation maybe this would be a good starting point for you to continue:

main...felixdittrich92:doctr:viptr-torch

Some points which are missing:

• 1. Cleanup layers
• 2. I had in mind to refactor the VIPBlock that the final VIPNet can inerhit from nn.Sequential comparable to all the other classification models we have -> avoid the mixer_types condition
• 3. Implement the recognition part -> VIPNet as feature extractor from the classification module + custom (linear) head + CTC loss + postprocessor + base class _VIPTR for target building
• 4. Testing everything (dummy run -> I can provide a dataset for testing or test it on my machine)
• 5. Add missing unittest + documentation entries

• port code to TensorFlow (optional - sec PR)

Maybe the vitstr PR as reference it shows all the required parts:

https://github.com/mindee/doctr/pull/1055/files

[felixdittrich92]

If there is any support needed feel free to ping or ref in a PR 👍

[lkosh]

I have refactored VIPBlock and VIPNet (5f936ea#diff-5d0b7beab5874f39f89df0bf5401c83ff4a991e2c22864a5768c4f179884933e) , now working on the recognition part