Training result of UETSC dataset differs from the pertained model about actor HOT 4 CLOSED

Hello jinseokbae,

This is indeed very strange. Did you download the already preprocessed data from this repo?
I tried to reproduce your failure without any success. I just retrain the model using Pytorch 1.10, but it works fine. What version of pytorch/cuda are you using?

Did you try again and got the same results?

Unfortunately, I am not sure of how I can help.

from actor.

Hi @Mathux,
I appreciate your reply.
I tried Pytorch 1.8.1 with CUDA version 10.1.
Also I downloaded preprocessed UESTC data as instruction guided.
If preprocessed dataset I downloaded is exactly identical to the dataset you used, Pytorch version might be an issue.
By the way, thank you for the current test on it :)

from actor.

Hi @jinseokbae ,

I encountered the same problem as you. I ran 1500 epochs and the results I got is
gen | 103.23 +/- 2.59 | 95.37 +/- 0.79 | 73.2 +/- 0.27 | 27.38 +/- 0.43 | 14.04 +/- 0.07
And the training loss I got is
| kl | 312.11390 |
| rc | 0.00760 |
| rcxyz | 0.00706 |
| mixed | 0.01778 |

I use PyTorch 1.11.0 with CUDA 11.6.
May I know your reproduced results?

Thanks!

from actor.

Hi @jinseokbae ,

I encountered the same problem as you. I ran 1500 epochs and the results I got is gen | 103.23 +/- 2.59 | 95.37 +/- 0.79 | 73.2 +/- 0.27 | 27.38 +/- 0.43 | 14.04 +/- 0.07 And the training loss I got is | kl | 312.11390 | | rc | 0.00760 | | rcxyz | 0.00706 | | mixed | 0.01778 |

I use PyTorch 1.11.0 with CUDA 11.6. May I know your reproduced results?

Thanks!

After several trials, I solved this problem by increasing learning rate, training epochs and reducing the num_layers to 4. Hope this can solve your problem.

from actor.