About mask2 in cifp_margin

[yiminglin-ai]

Hello, why in CIFP margin did you consider only neg_logits that are smaller than pos_logits - self.m? Why those largerneg_logits should not be penalised? Thanks.

MixFairFace/network.py

Lines 29 to 30 in 11691d3

	mask2 = (pos_logits - self.m) > neg_logits
	topk_mask = topk_mask * mask2

[fuenwang]

This part is modified from official CIFP code (https://github.com/Tencent/TFace/blob/99cc558fde0176f38ff051155b110bfd7f6cdb3a/recognition/torchkit/head/localfc/cifp.py#L12)

I don't know the reason either. But I have tried many experiments on CIFP, and the implementation can really work and improve accuracy.

[yiminglin-ai]

Thanks Fu-En, your implementation is much more readable than the official code! Have you experimented without the mask2?

[fuenwang]

I guess so. Because I had removed the official code line-by-line to check the corresponding performance, and my network.py only keeps the necessary parts to improve performance.

[yiminglin-ai]

I see. My assumption is that the high neg_logits may be due to label errors, such as images with incorrect ID labels. Therefore, CIFP avoids using these neg_logits exclude these errors. They can be harmful especially towards the end of the training process.