Fail to build val dataloader about keypointnerf

[markomih]

Hi Fuqianya, thank you for reporting the issue!

We use only these three subjects (CoreView_387, CoreView_393, CoreView_394) for evaluation.

Have you by chance selected one of the training subjects in the val/test data loader?

[fuqianya]

Oh, yes. Thanks for your quick reply!

It's my mistake. Since I only download the CoreView313, so I evaluate on it. This issue can be fixed by not performing evaluation on the CoreView313 and 315.

I have another problem: How many gpu times are used for training on the ZJU-MOCAP? And which gpu are your used? I had tried on my 3090, but it failed due to the limited gpu memory.

[markomih]

We used V100s to train the models. You could reduce the patch size here for lower memory usage.

The validation/rendering script can run on a 24GB GPU.

[fuqianya]

All right, thanks~

I have evaluate the pretrained model on my 3090 successfully..

How about the training time? How many gpus are you used for training?

[markomih]

We used only 1 GPU.
The multi-gpu training is supported as well.

[fuqianya]

ok, thanks for your rely! I will close this issue.