Jianyuan Wang comments

Results 238 comments of


                                            Jianyuan Wang

TRAINING VGGT fails with CUDA OOM

Hey 8GB would be too small for training.

Questions about training stability

Hi, 1. Yes, DPT can indeed produce very large gradients 2. Based on discussions with others working on similar pipelines, successful training seems to depend heavily on well-tuned hyperparameters. Some...

Questions about training stability

I used check_and_fix_inf_nan() mostly to detect nan/inf. It can help but cannot totally solve the problem.

Reconstruction has missing attributes

Hi, you can reduce the confidence to around 10 or 20, and usually those missing parts will appear

Loss converged upon unfreezing aggregator but results bad

Hi can you try to use a smaller lr, e.g., 10% of the original one?

Some questions about ba

Hi, I am a little confused about this question. Could you explain it in detail?

BA

Hi we have not provided the implementation of BA in this repo, but you can check out [VGGSfM](https://github.com/facebookresearch/vggsfm) for an early look

BA

Closing this issue as BA implementation has been uploaded.

Co3d Annotation files

Hi, I have uploaded it here: https://huggingface.co/datasets/JianyuanWang/co3d_anno/tree/main

About total loss for fine-tuning on custom dataset

Yeah this is converging relatively slow. Ideally with so many steps the avg loss_R should be around 0.01.