Should we train the prior model with n_cond_frames>=1 for BAIR Pushing dataset?

wilson1yan / VideoGPT

MIT License

962 stars 115 forks source link

Should we train the prior model with n_cond_frames>=1 for BAIR Pushing dataset? #12

Closed xinbowu2 closed 3 years ago

xinbowu2 commented 3 years ago

I noticed the generated samples shown on the paper for BAIR Pushing dataset are conditioned on a single frame, so I am wondering if we should also train the prior model conditioning on a single frame. However, I saw the default setting for n_cond_frames is 0 for BAIR Pushing dataset from another repo: VideoGPT-Paper. Also, I'd like to know if the comparison with other methods on BAIR Pushing Dataset is also based on single-frame conditioning.

wilson1yan commented 3 years ago

Yes, for purposes of fair comparisons with prior work, you should train with --n_cond_frames 1. All tables / results in the paper related to BAIR are under the single-conditional setting.