mccreezhao / qamface Goto Github PK

Ideally the input data is normalised between -1 and 1 because fp32 and fp16 have better precision in that range.
Is there a reason why it is commented out in the code?

the loss function used in the project was loss = self.focalLoss(thetas, labels)??? not Additive Angular Margin Loss?

Hello and thanks for your work!

Did you try to use cross-entropy loss with your QAAM head? If yes, what were the results compared to the focal-loss?

I notice you only went to epoch 22, which is still stage 1. Are you still training this?

The batch size in config.py is not used for anything (instead it used the -b option).
This confused me initially. Maybe it could change the default value.

I haven't seen your paper yet but was just wondering how it was determined to use an initial LR of 0.2 instead of the typical 0.1.
While training IR_152 backend, I saw training loss jump significantly to 100.
Also, have you tried retraining using a pretrained backbone?

I noticed there's no code to set which GPUs are to be used. There's also no synchronisation code in the loss function for multiple GPUs.

How did you train it on 4 GPUs without this, or is it no longer needed in modern PyTorch?

I modified your QAMFace to include the support:

        if self.device_id == None:
            kernel_norm = l2_norm(self.weight, axis=0)
            cos_theta = torch.mm(embbedings, kernel_norm)
        else:
            x = embbedings
            sub_kernels = torch.chunk(self.weight, len(self.device_id), dim=1)
            temp_x = x.cuda(self.device_id[0])
            kernel_norm = l2_norm(sub_kernels[0], axis = 0).cuda(self.device_id[0])
            cos_theta = torch.mm(temp_x, kernel_norm)
            for i in range(1, len(self.device_id)):
                temp_x = x.cuda(self.device_id[i])
                kernel_norm = l2_norm(sub_kernels[i], axis = 0).cuda(self.device_id[i])
                cos_theta = torch.cat((cos_theta, torch.mm(temp_x, kernel_norm).cuda(self.device_id[0])), dim=1)