denormalize is missing in eval_long

[Kairobo]

No description provided.

[kleinzcy]

Same issue.

When you have time to open-source the evaluation of long-term generation?

[Sirui-Xu]

Hello,

Here’s the code for denormalizing and evaluation: link to code. It’s not fully organized yet but should produce the same results as reported in the paper.

Will find a time to complete the wrap-up for this part and integrate it into the current codebase.

Best

[kleinzcy]

Thanks for your code. The provided evaluation code contains the denormalize function. But the evaluation is still failed as it missing "correct" function (Line 314 in eval_smpl_long.py).

[kleinzcy]

Another question. It takes 10 minutes for 1 epoch on the TITAN RTX. Is it normal? And it occures oserror when num_workers > 0. How do you fix such error?

[Sirui-Xu]

The ‘correct’ function should be ‘smooth’ as defined in the short-term evaluation code, sorry for the confusion

I don’t recall encountering similar cases before in a Linux environment with A40 or TITAN X. The network and dataset are lightweight and should train very quickly.

[kleinzcy]

Thanks for your help. However, the evaluation still failed as the get_batch function is not right. The repeat dimension is not right. I have tried, but not solved.