Do we need to scale word embeddings to [-1, 1]? about diffusion-lm

[XiangLi1999]

Hi,

Thanks for the question. We are not mapping the word embeddings to be between [-1, 1], and this is different from image diffusions.

There are three terms in the objective: (1) Lsimple (mse), (2) the reconstruction (i.e. decoder_nll) (3) the prior (t_T_loss) as in

Diffusion-LM/improved-diffusion/improved_diffusion/gaussian_diffusion.py

Line 1577 in 759889d

terms["loss"] = terms["mse"] + (decoder_nll + tT_loss)

. Term (2) prevents norm from being too small, term (3) prevents the norm from being too large.

Hope this helps!

[tj-zhu]

Yes this explains it! Thank you very much for the quick response and the great explanation!

[tj-zhu]

Hi @XiangLi1999, I am sorry for reopening the issue. I just have one more question about the loss function.

Can I ask why in decoder_nll loss, we input x_start instead of the predicted x_start?
You mentioned the decoder_nll is to prevent word embeddings being too small. I assume it's because if the word embeddings are too small, the noise will take dominance, and it will be difficult for model to denoise, then the reconstruction loss will be high? Please correct me if I am wrong.

If that's the purpose of this reconstruction loss, then we need to use the predicted x_start (the denoised version) to calculate reconstruction loss, right?

Sorry if the answer seems obvious but I didn't get it. Thank you very much for your help!

Diffusion-LM/improved-diffusion/improved_diffusion/gaussian_diffusion.py

Line 1569 in 759889d

decoder_nll = self.token_discrete_loss(x_start, get_logits, input_ids)