search

IshmaelBelghazi / ALI

Adversarially Learned Inference
MIT License
311 stars 80 forks source link

mistake in D(x,z) input size #10

Closed edgarriba closed 6 years ago

edgarriba commented 7 years ago

In table 5 from the paper you state that the input size for D(x,z) is 1024x1x1 which I think it's wrong after looking at the previous output sizes D(x) and D(z). I think that should be 1536x1x1.

Is that assumption correct?

edgarriba commented 7 years ago

ping @IshmaelBelghazi @vdumoulin

vdumoulin commented 7 years ago

Sorry for the delay; I'll have a short look at get back at you on this.

vdumoulin commented 7 years ago

@edgarriba You are absolutely correct, thank you for pointing this out!

hubert0527 commented 6 years ago

I have a similar problem about the hyper-parameters in Table 5.

The output feature map channel of G_z(x) (i.e. 512) is supposed to be 2 times the size of the latent vector dimension (due to the VAE architecture). But G_x(z) shows that the latent vector size is 512, instead of 256. I'm not sure which one is wrong? Or I misunderstand something? Thanks!

vdumoulin commented 6 years ago

@hubert0527 comparing Table 5 in the paper and the relevant portion of the source code, I believe there is a typo in Table 5: G_x(z) – 512 × 1 × 1 input should instead read G_x(z) – 256 × 1 × 1 input.

Thanks for pointing it out, and sorry about the confusion!