Train ArcFace on VGGFace2 dataset

[ThanosRoidis]

I am trying to train a ResNet-34 on the VGGFace2 dataset using ArcFace loss (custom Keras implementation), but I can not get better results than regular Softmax Cross-Entropy loss.

In the published ArcFace paper, on Table 7, you report the results of a (VGG2, R50, ArcFace) model which I am trying to replicate, but you do not specify anywhere the paper the training arguments / hyperparameters when training on VGG2, other than m=0.3, and I can not find them anywhere in this repo.

Can anyone please tell me where can I find them (learning rate, steps, dropout, preprocessing, etc)

[nttstar]

For vgg2, just set m=.3 and lr drops at 40K, 60K, 80K iterations.

[ThanosRoidis]

@nttstar Starting at 0.1 and then divided by 10 with a batch size of 512 (momentum 0.9) as noted in the paper for the other datasets?

Also, the input is 112x112 (mtcnn aligned) and the embedding size 512, or 224x224 and 2048 as in the vgg2 paper?

[nttstar]

These are our base settings: batchsize: 512 ; 112x112 aligned; emb-size 512.

[ThanosRoidis]

Final question, on what iteration do you stop training?

[jacksoncsy]

Dear both, I have the same question, when do you stop training? at which epoch/iteration? Many thanks!

[nttstar]

@jacksoncsy Anytime after 80K iters.

[jacksoncsy]

Thanks Jia! May I further ask you how to convert your iterations to epochs? For example, in the setting of 8x GPUs, batch size 128 per GPU, 3m VGGFace2 images, 80K iters equals to how many epochs? Many thanks!

[nttstar]

I haven't trained vgg2 for quite a while. The total batch-size is 512 and you can easily estimate the epoch numbers.

[jacksoncsy]

Thanks, I see! So essentially, not many epochs for vggface2 then.