About the "pixpro_instance_loss" question ???

[LiShuiYu]

Thank you very much for your work and code sharing. In the Pixpro.py, I found that in "pixpro_instance_loss", 'pred_instance_1' has one more predictor layer than 'proj_instance_1_ng'. I don’t quite understand why this is necessary. I am waiting for your reply. Thank you very much.

   if self.pixpro_ins_loss_weight > 0.:
        proj_instance_1 = self.projector_instance(feat_1)
        pred_instacne_1 = self.predictor(proj_instance_1)
        pred_instance_1 = F.normalize(self.avgpool(pred_instacne_1).view(pred_instacne_1.size(0), -1), dim=1)

        proj_instance_2 = self.projector_instance(feat_2)
        pred_instance_2 = self.predictor(proj_instance_2)
        pred_instance_2 = F.normalize(self.avgpool(pred_instance_2).view(pred_instance_2.size(0), -1), dim=1)

    # compute key features
    with torch.no_grad():  # no gradient to keys
        self._momentum_update_key_encoder()  # update the key encoder

        feat_1_ng = self.encoder_k(im_1)  # keys: NxC
        proj_1_ng = self.projector_k(feat_1_ng)
        proj_1_ng = F.normalize(proj_1_ng, dim=1)

        feat_2_ng = self.encoder_k(im_2)
        proj_2_ng = self.projector_k(feat_2_ng)
        proj_2_ng = F.normalize(proj_2_ng, dim=1)

        if self.pixpro_ins_loss_weight > 0.:
            proj_instance_1_ng = self.projector_instance_k(feat_1_ng)
            proj_instance_1_ng = F.normalize(self.avgpool(proj_instance_1_ng).view(proj_instance_1_ng.size(0), -1),
                                             dim=1)

            proj_instance_2_ng = self.projector_instance_k(feat_2_ng)
            proj_instance_2_ng = F.normalize(self.avgpool(proj_instance_2_ng).view(proj_instance_2_ng.size(0), -1),
                                             dim=1)

[LiYuhangUSTC]

I think it is also an asymmetric architecture design, similar to PPM. See BYOL paper (ref [18] in PixPro paper: https://proceedings.neurips.cc/paper/2020/file/f3ada80d5c4ee70142b17b8192b2958e-Paper.pdf) for details.