Hi,Dr.Zhang,Thanks for your paper(ZERO-CL) and your hard work.

[LYJhere]

I have some question about the following codes { def whitened_barlow_twins_loss(self, y1, y2): z1 = self.whiten_bn(self.projector(self.backbone(y1))) z2 = self.whiten_bn(self.projector(self.backbone(y2)))

    if self.args.whiten=='true':
        z1 = self.whiten_net.zca_forward(z1)    # N * d
        z2 = self.whiten_net.zca_forward(z2)

    c = torch.mm(z1.transpose(0, 1), z2)         #d * d
    c.div_(self.args.batch_size)
    loss = torch.diagonal(c).add_(-1).pow_(2).sum()
    if self.args.off == 'true':
        off_diag = off_diagonal(c).pow_(2).sum()
        loss += self.args.lambd * off_diag
    return loss

} Can you answer me the following question: 1、You can calculate the loss about the instance and feature level. But I only find the loss about feature level (d * d).

[Sherrylone]

You can add z1 = z1.transpose(1, 2) before z1 = self.whiten_net.zca_forward(z1). Then, the whitening process will be calculated on the instance dimension.

[LYJhere]

Yes , thanks for your answer.