Finetuning ArcFace for different domain

[AlexanderHustinx]

First off, thank you very much for this repo. It has been very insightful and performs nicely.

Now, I have a problem where I'd like to use face recognition as a base for transfer learning a different task. As such I'd like to use the ArcFace weights and finetune them on my own task, updating the representations in the process.

However, I am not sure how to do this using the ONNX model weights you have provided. Could anyone please shed some light on that for me?

Thanks in advance!

[43ig2]

Hi,

I encounter the same problem here. A way to do transfer learning from pretrained ONNX models would be to convert them into Torch format. However I could not find a good way to do it. Did someone find any solution to do the ONNX -> Torch model weights conversion for the iresnet50-based models? If this is not possible, it would be nice to consider releasing trained models in .pt format in addition to the .onnx format.

Thanks for your help.

[gerald-ftk]

The repo is a bit messy, but the authors have actually uploaded Pytorch models here: https://onedrive.live.com/?authkey=%21AFZjr283nwZHqbA&id=4A83B6B633B029CC%215577&cid=4A83B6B633B029CC

[brtang63]

The repo is a bit messy, but the authors have actually uploaded Pytorch models here: https://onedrive.live.com/?authkey=%21AFZjr283nwZHqbA&id=4A83B6B633B029CC%215577&cid=4A83B6B633B029CC

Thanks a lot. I am new to pytorch and haven't figure out how to use the models here. Does the directory contain the full model or just the models weights? By the way, may I ask where do you find this link? Is there any illustrations? @gerald-ftk

[AlexanderHustinx]

@43ig2, I ended up using onnx2torch from https://github.com/ENOT-AutoDL/onnx2torch Basically this snippet should do it:

from onnx2torch import convert

base_model = "saved_models/glint360k_r50.onnx"    # use your own model and path
self.base = convert(base_model).to(device)

and then just add your own layers on top as required and as you normally would, e.g.:

self.classifier = self.classifier = nn.Sequential(nn.ReLU(True), nn.Dropout(), nn.Linear(512, num_classes))

and in the forward-function just, as normally:

def forward(self, x):
    x = self.base(x)
    y = self.classifier(x)
    return y

Small caveat, it doesn't seem to work for all models. You might also need to make sure the ArcFace-class is available somewhere.. I'm not 100% whether that was necessary anymore, it's been a while.

@gerald-ftk, I was aware of these models (thanks though). But there are many more models they provide a *.ONNX-file here: https://github.com/deepinsight/insightface/tree/master/model_zoo The snippet I provided above can be used with some success there.