[YOLOv8] Step 0/? Addition of SPP and SPPF Layer

[IMvision12]

Added Spatial Pyramid Pooling (SPP) and SPPF layer.

Fixes: #1263
SPP paper : https://arxiv.org/abs/1406.4729
SPPF layer was created by Glenn Jocher which is an optimized version of SPP.

Question: should I create series of many PRs for adding layer 1 by 1 or single PR for everything?

@ianstenbit @LukeWood

[ianstenbit]

Hi @IMvision12 -- super excited about this!

Do you have a full version of YOLOv8 working, or are you planning to build it out piece by piece?

For code review, it would be nice to have small PRs for one component at a time, but it'd be great to have a full working Colab before we start merging pieces individually.

[IMvision12]

@ianstenbit Yeah I am working on colab, will share it

[LukeWood]

Thanks for the PR! I left a first batch of changes on the PR.

[IMvision12]

@LukeWood @ianstenbit are we going to add a training script too? till now I have completed with model and loss currently working on training!!!

[ianstenbit]

@LukeWood @ianstenbit are we going to add a training script too? till now I have completed with model and loss currently working on training!!!

Hi @IMvision12 -- this is exciting!
A training script would be great, yes. Ideally this would start as a Colab that we can later add as an example in KerasCV.

In order to verify correctness, we should find a reference implementation and see if we can port its weights and verify numerical equivalence.

[haifeng-jin]

Hi @IMvision12 , thank you so much for the contribution!
We are prioritizing the YOLO model in KerasCV.
I will be assisting you in merging your PR.

Would you please update us with the current status of this PR?

[IMvision12]

Hii, @haifeng-jin, This PR is probably ready to merge, but as @ianstenbit proposed, we need to first build a complete functional model before we can start merging PRs one by one, so I'm working on creating a full implementation of yolov8 here on my repo: https://github.com/IMvision12/YOLOv8-tf, currently I am working on loss, modeling part is completed.

[haifeng-jin]

Thanks @IMvision12 ! Looking forward to it.

[haifeng-jin]

Hi @IMvision12, would you please let us know a estimated timeline?
We would like to coordinate this line of work with our overall release schedule of KerasCV.
Thank you!

[IMvision12]

Hello, @haifeng-jin, YOLOV8, and other OD and segmentation models align with my GSoC timeline, so if possible can you please go through the timeline in the proposal? There's a detailed timeline on the proposal.

Short answer: I can finish Yolov8 until the last week of April because I have university examinations, and then I can start submitting PRs the first week of May, does this align with your timeline for the release of KerasCV?

[jbischof]

@IMvision12 thanks so much for your contribution! We haven't accepted any GSoC proposals yet and will only have one slot for KerasCV. Would you still be interested in contributing this model outside the GSoC program?

[IMvision12]

Yes @jbischof

[IMvision12]

@ianstenbit @haifeng-jin @jbischof @LukeWood I can start sending PRs from first week of may

[IMvision12]

#1675