questions on source data

[louisefz]

Hi summmeer,

I have one question on dataset of source side. You mentioned in your paper that the source part serves as condition for the target part, and noises will only be imposed on target part. However, I didn't see such conditioning. Besides, how to process source together with the noising target during the forward process; I guess the final zT can be concat(Emd(x),yT). In the reverse process or smapling stage, how will source part be functioning?

[summmeer]

Hi,
The $x$ in our code corresponds to $z$ in the paper, and we use mask to impose the partial noise. You can refer to:

DiffuSeq/diffuseq/gaussian_diffusion.py

Line 400 in f78945d

sample = th.where(mask==0, x_start, sample)

DiffuSeq/diffuseq/gaussian_diffusion.py

Line 259 in f78945d

return th.where(mask==0, x_start, x_t)