issue about reproducing results on SNLI dataset

[20000607-lxc]

Hi! I'm trying to reproduce the reported result on SNLI, I followed the doc 'optimus_for_snli.md' and successfully downloaded the checkpoints, but when I run your examples, it turns out that in file run_latent_generation.py, the sample_sequence_conditional function receives 'input_ids' and 'past' in mismatched shape. I can fix this by past = torch.mean(past, dim=0).unsqueeze(0), but is it right? Thanks for reading.

[20000607-lxc]

Hi! Another problem is that when running 'run_lm_vae_training.py' with Snli data you provided, I got extremely large perplexity = 4.017335065391466e+66, and kl = 1922.1040397033692.