a bug about training the full pipeline 1024x1024

[yjh576]

Hi! I have a problem about training the full ffhq pipeline 1024x1024.

[royorel]

Hi @yjh576,

Can you please specify the exact setup and hyper-parameters you used?

I can't reproduce the bug on my end.

[yjh576]

I downloaded it directly from your codebase and did not make any changes.

[royorel]

I mean what command did you run? and which arguments did you use?

[royorel]

@yjh576 I was able to reproduce this issue now, I'm still not sure about the cause, but this may potentially require a fix for the code. I will update once I will resolve this.

Thanks for reporting this issue!!

[royorel]

Hi @yjh576,

What pytorch version are you using? This issue seems to rise if pytorch version > 1.9. When I used pytorch 1.10.0 & 1.11.0 the bug happened, but when I used pytorch 1.9.0 the code ran smoothly.

I still don't know the reason it happens, but the best workaround is to run the code with pytorch 1.9.0.

Let me know if there are still any issues

[yjh576]

Pytorch version I used is 1.10.0. Thank you for your help.

[Ni1r]

@yjh576 hi ,i met an error when training the volume renderer,Can I take a look at the structure of your dataset ,please?thanks!

[royorel]

Hi @Ni1r,

It would help if you specify the actual error. As for the dataset, before running prepare_data.py the dataset should only contain image files. In addition, if you made it through the volume renderer training phase with no issues (I'm assuming you ran that first), then the error is not in your dataset structure.

[Ni1r]

Hi @Ni1r,

It would help if you specify the actual error. As for the dataset, before running prepare_data.py the dataset should only contain image files. In addition, if you made it through the volume renderer training phase with no issues (I'm assuming you ran that first), then the error is not in your dataset structure.

this is my error,before this ,i have run the "prepare_data.py" and get the data.mdb and lock.mdb