about output loss definition

[hinanmu]

Hi,I have a question in your code about loss definition. I notice that the prediction of the output loss in the C2AE paper is defined as F_d(F_e(x_i)). But it is defined as F_d(F_x(x_i)) in your code.

Is it a mistake in the paper? I don't know why F_e function can accept x_i, they have a different dimension. Thank you!

[greeness]

I think there is a mistake in the paper.

In the training time, to get the loss between Fe and Fd, we need to calculate Fd(Fe(y_i)) (Not x_i, because as you pointed out, Fe cannot take x_i as input but y_i).

Probably because of this mistake, dhruvramani@ used a different way to derive the loss. It is probably working too but not as good as the one proposed by the original paper.

In the prediction time, to predict labels, we would do

Fd(Fx(x_i)) this is consistent between the formula from the paper and the implementation here.

[greeness]

I found out that in the latest version of the paper hosted on arXiV . That mistake is already fixed. Now equation (3) cotains Fd(Fe(y_i)) instead of Fd(Fe(x_i)).

See https://arxiv.org/pdf/1707.00418.pdf

[hinanmu]

Thanks, I got it.

[chihkuanyeh]

@greeness thanks for the clarification.