Add MAXIM model

[innat]

Short Description

MAXIM: Multi-Axis MLP for Image Processing: I think it's a follow-up work of MaxViT from google. It shows a great performance on the following low-vision task, i.e. for OCR.

It looks useful for real world problem (blurry, haze, rain-drop). Such problems are very acute for vision project. So, such solutions might weight quick adaptation of kcv in industrial projects (alike).

Papers

• https://arxiv.org/pdf/2201.02973.pdf

Existing Implementations

• In Jax: https://github.com/google-research/maxim
• maxim-tf
• Google-blog: https://ai.googleblog.com/2022/09/a-multi-axis-approach-for-vision.html
• pytorch: https://github.com/vztu/maxim-pytorch (WIP).

cc. @vztu

[tanzhenyu]

Open this up for contribution

[innat]

cc @sayakpaul

[IMvision12]

I'm working on integrating the maxim model and referencing the original flax implementation. I have one question: do we need weights as well? @tanzhenyu

[tanzhenyu]

I'm working on integrating the maxim model and referencing the original flax implementation. I have one question: do we need weights as well? @tanzhenyu

Yes, we need the weights. The development cycle would be similar to ViT: https://github.com/keras-team/keras-cv/pull/1014, add the model and port the weights, also add the conversion scripts for porting the weights

[bhack]

Yes, we need the weights. The development cycle would be similar to ViT: https://github.com/keras-team/keras-cv/pull/1014, add the model and port the weights, also add the conversion scripts for porting the weights

I hope we could standardize a bit the model contribution process after we will handle:

https://github.com/keras-team/keras-cv/discussions/954

[IMvision12]

I'm working on integrating the maxim model and referencing the original flax implementation. I have one question: do we need weights as well? @tanzhenyu

Yes, we need the weights. The development cycle would be similar to ViT: #1014, add the model and port the weights, also add the conversion scripts for porting the weights

Okay will add the conversion script too!!

[innat]

@IMvision12 Have you checked ? https://github.com/sayakpaul/maxim-tf

[IMvision12]

@IMvision12 Have you checked ? https://github.com/sayakpaul/maxim-tf

Oh we have tf implementation for maxim, because when I saw on paper with code there was only official flax implementation will refer that too thx for the info @innat. Also as we already have weights ported to TF and uploaded by @sayakpaul on TF-Hub do we still require the conversion script? @tanzhenyu