Validation accuracy & Mobilenet

[JohannesTK]

Hi,

Thank you for the rewrite and very intuitive repo.

Have you reported validation accuracy for the Resnet50 and have you also implemented mobilenet?

Thanks, Johannes

[ternaus]

I did not.

It is not that hard to add the whole set of backbones from https://github.com/rwightman/pytorch-image-models/blob/master/results/results-imagenet.csv but I do not really need it right now, hence I cannot tell when this functionality will be added.

[JohannesTK]

Thanks for a quick response and well understood regards mobilenet.

Regards the Resnet50 val accuracy, is it the same as reported in https://github.com/biubug6/Pytorch_Retinaface#widerface-val-performance-in-single-scale-when-using-resnet50-as-backbone-net or do you have some numbers for that?

[tloki]

I doubt it is the same since @ternaus did train his own model.

A validation accuracy (and other numbers) would be nice just to compare to the baseline retinaface.

[woctezuma]

For information, there is also another project called FaceXLib with RetinaFace for detection.

The Python package is available on PyPI:

pip install facexlib

As shown in the following code:

https://github.com/xinntao/facexlib/blob/45fee8e7885fc462d2d4c4eafb2650ff384b89c8/facexlib/detection/init.py#L8-L16

The model checkpoints are downloaded from here:

https://github.com/xinntao/facexlib/releases/tag/v0.1.0

They use the same name convention as the original model checkpoints, so I wonder if these are direct copies without re-training. Edit: See my following post, the answer is yes.

https://github.com/biubug6/Pytorch_Retinaface#training

[woctezuma]

I have run the following code on a Google Colab machine:

!wget https://github.com/xinntao/facexlib/releases/download/v0.1.0/detection_mobilenet0.25_Final.pth
!wget https://github.com/xinntao/facexlib/releases/download/v0.1.0/detection_Resnet50_Final.pth

!gdown --id 15zP8BP-5IvWXWZoYTNdvUJUiBqZ1hxu1
!gdown --id 14KX6VqF69MdSPk3Tr9PlDYbq7ArpdNUW

!sha256sum *.pth

This is the output:

2979b33ffafda5d74b6948cd7a5b9a7a62f62b949cef24e95fd15d2883a65220  detection_mobilenet0.25_Final.pth
6d1de9c2944f2ccddca5f5e010ea5ae64a39845a86311af6fdf30841b0a5a16d  detection_Resnet50_Final.pth
2979b33ffafda5d74b6948cd7a5b9a7a62f62b949cef24e95fd15d2883a65220  mobilenet0.25_Final.pth
6d1de9c2944f2ccddca5f5e010ea5ae64a39845a86311af6fdf30841b0a5a16d  Resnet50_Final.pth

So the model checkpoints should be the same.

--

If you don't have access to sha256sum on your system (maybe Windows), you can do it in Python:

# Reference: https://stackoverflow.com/a/44873382/376454

import hashlib

def sha256sum(filename):
    h  = hashlib.sha256()
    b  = bytearray(128*1024)
    mv = memoryview(b)
    with open(filename, 'rb', buffering=0) as f:
        for n in iter(lambda : f.readinto(mv), 0):
            h.update(mv[:n])
    return h.hexdigest()

import glob

for fname in glob.glob('*.pth'):
  print(f'{sha256sum(fname)} {fname}')

This is the output (same as above):

2979b33ffafda5d74b6948cd7a5b9a7a62f62b949cef24e95fd15d2883a65220 mobilenet0.25_Final.pth
6d1de9c2944f2ccddca5f5e010ea5ae64a39845a86311af6fdf30841b0a5a16d detection_Resnet50_Final.pth
2979b33ffafda5d74b6948cd7a5b9a7a62f62b949cef24e95fd15d2883a65220 detection_mobilenet0.25_Final.pth
6d1de9c2944f2ccddca5f5e010ea5ae64a39845a86311af6fdf30841b0a5a16d Resnet50_Final.pth

[tloki]

How does this have anything to do with this repo? If I understood you right - your claim is that FaceXLib uses same weights as biubug6 (original retinaface repo?). On FaceXLib it is even stated (README) that they do use biubug6 model/code...

[woctezuma]

How does this have anything to do with this repo?

It is additional information for users like myself who:

• have used this repository in the past,
• would like to have validation accuracy for the model checkpoints which they use.

FaceXLib is an alternative which:

• did not exist when this repository was created,
• is installable with pip and just as easy to use as this repository,
• relies on the original weights, for which accuracy is documented.

[tloki]

The reason I was actually interested in this repo was - NMS. biubug6 uses its own implementation, which is slow (it could somewhat easily be parallelized), and this one uses off-the shelf NMS which seems like a less of bottleneck. I'm mainly refereing to for loop which does NMS for every element in batch.

[woctezuma]

I see. It looks like this repository relies on:

• torchvision.ops.nms

wile FaceXLib relies on:

• a custom Python implementation.

I'm mainly referring to for loop which does NMS for every element in batch.

Maybe I misunderstood and you are referring to another part of the code than the one which I mentioned.