Implementing CFM and L2BT

[alex-costanzino]

What is the motivation for this task?

Hey everyone, not sure I'm using this correctly, so forgive me in advance :)

I would like to propose the inclusion of two of my projects, CFM and L2BT, into the Anomalib library. Both have demonstrated remarkable performance in AD tasks while maintaining extremely fast inference speeds and a lightweight footprint.

These projects are versatile and cover different modalities (multimodal and image-only), enhancing the diversity of models within Anomalib. Additionally, their underlying methodologies are quite distinct (feature mapping and feature transfer) from the current implementations in the library, which could broaden the scope of Anomalib.

Thank you for considering this request. I’m happy to provide further details or assist with integration.

Describe the solution you'd like

Is there any plan to integrate these approaches?

Additional context

L2BT also come with a protocol to fairly assess and evaluate AD methods, it would be interesting to integrate this as well somehow.

[samet-akcay]

Hi @alex-costanzino, thanks for your interest in Anomalib!

You may not recall, but you explained CFM to us at CVPR 24 during your poster session :)
I'm happy to have it in Anomalib. Let us know if you need anything from our side

[alex-costanzino]

Hi @samet-akcay, of course I remember! I also told you then that I was interested in integrating it, but then it took me a while to actually have some time to do that xD

Is there a guide or something that explains how to implement a method under Anomalib? My software engineering skills are a bit limited, but I could try to open a pull request.

[samet-akcay]

great! i'm currently working on documentation showing how to implement models in anomalib, but it's not ready yet.

You could potentially have a look at other models, but here is the overall idea:

• You need to implement a torch model and save it as torch_model.py, where forward method should return a loss value if the model is in train mode, and return your predictions in eval mode (such as anomaly map, scores etc).
• Second step is to implement the lightning module by inheriting AnomalibModel. This ensures the model performs the right action in its corresponding hook.
• If you have other custom modules such losses, anomaly map heads etc, you could define them as a separate module.

Once you finish your lightning_model.py, Anomalib should ideally take care of the rest.

You could have a look at this implementation as an example
https://github.com/open-edge-platform/anomalib/tree/main/src/anomalib/models/image/supersimplenet

If you feel stuck at any point, feel free to reach out. I'm more than happy to help.

Meanwhile, I'll try to finish the documentation