Proposal: PyTorch backend support

[elifkcontar]

Description

Hi,

DeepFace is tied to TensorFlow/Keras right now. For PyTorch-based projects this is a bit heavy since it means pulling in TF just for DeepFace, which can cause conflicts.

I’m considering working on a PyTorch backend:

• re-implement recognition models (ArcFace, VGG-Face, Facenet, etc.) in PyTorch
• keep the same API (verify, analyze) so users can switch between TF and Torch

Would you be open to this idea?

• As a PR to this repo (e.g. backend="torch")
• Or as a forked repo with clear credits to DeepFace

Curious what you think.

Additional Info

No response

[serengil]

IMO, having both tf and torch support makes repo heavier and hard to maintain.

Forking the repo is totally fine.

[serengil]

Wonder your opinions also.

[elifkcontar]

I'm think about two options:

If I go with a fork, I’d use existing PyTorch implementations for models where possible(e.g., retinaface-pytorch) and convert the rest via ONNX(if possible), then I’d run performance benchmarks and add comparisons to make it transparent for the community.

If I go with a PR, I’d introduce modular backend support (similar to what Keras used to do in its early days). The API would stay the same, and Torch would only install as a dependency if backend="torch" is used, so TF users wouldn’t be affected.

Curious to hear your thoughts too.

[serengil]

i am thinking those pros and cons

• if we put torch related functionalities in original deepface repo, when you perform pip install deepface then it will install tensorflow dependencies by default. users may need to run pip install deepface --no-deps to avoid install tensorflow, and install other dependencies manually. standard users will be able to use deepface with tf, but users need torch have to manage these. not too bad. we also need to write another ci cd pipeline for torch related tests.

• if we seperate deepface and deepface-torch repos, i strongly recommend to give a different import name because one may have both deepface and deepface-torch, and from deepface import DeepFace will cause conflicts. We should have something like from deepfacetorch import DeepFace.

• having seperate repos may cause incompabilities for new features. I mean that new features can be added to deepface, but they may be missing in deepface-torch.

[FurkanGozukara]

this is literally mandatory for this project to be useful on Windows

#1512