This repository contains the data and code for the paper "Diverse Text Generation via Variational Encoder-Decoder Models with Gaussian Process Priors" (SPNLP@ACL2022)
When I am trying to train the t5-gpave, there is a problem of "zero kld!!!".
Also, the "zero kld!!!" is also in the training of the LSTM-based variational encoder-decoder with GP priors.
Thank you for your help and I am looking forward to hearing from you.
Hi!
Thanks for your great work! I'm working on getting results on another paraphrase dataset under T5 + GP prior setting. I have the following two questions:
I found that the generation speed is relatively slow due to the inference batch size 1, and something get wrong if I change it. Is there any way to speed up the generation?
if I want to get a trade-off between quality and diversity, is it suitable to set the scalar to 7 just like it used in the paper for the paraphrasing task?
Thank you for your great work! I am learning VAE recently, your paper have given me great inspiration.
In naive VAE, the KL divergence should be $KL(\mathcal{N}(\mu, \sigma), \mathcal{N}(0,1))$. But when reading your code, I found model/t5/t5_vae.py line 151, you set the prior_logvar as 0. Is there any mistake?