Jianyuan Wang comments

Results 238 comments of


                                            Jianyuan Wang

Any workarounds for improving speed on V100 gpus?

I see, this seems better but still quite slow. Unfortunately I do not have access to V100 so not sure how to accelerate it, but I guess someone has tried...

Preprocessing details for Replica

Hi, I think 512 is the original resolution. You do not need to upsample the images. The resizing is automatically solved within the dataloader, as in https://github.com/facebookresearch/vggt/blob/44b3afbd1869d8bde4894dd8ea1e293112dd5eba/training/data/base_dataset.py#L115

Bug In trainer.py while accum_steps ＞ 1

Hi here is the response from a llm (I did not carefully check it though). ``` Based on the code provided, the comment/bug report is incorrect. The logic in get_chunk_from_data...

Applying texture to 3D point map output

Hi, if your texture has a 1-1 mapping to the depth map, it should fine to just use them?

RTX5070Ti. struggled with the torch version and CUDA version

Hi! It looks like you’re running on a Windows machine. Installation on Windows needs some manual customization, as our provided script can’t handle it as is.

The problem of incorrect scale scaling in the 3D modeling

Hi our prediction is in a normalized space

Update demo_gradio.py

Hi would you mind talking more about the goal of adding an extra reload button? It seems the visualisation will be auto-reloaded once some settings are changed.

Update demo_gradio.py

I see! That makes sense. Let me think for a while how should this be designed.

RuntimeError: batch size must be positive when accum_steps > batch size in dynamic batching

Yes, this is expected. If you want to accumulate N steps, the batch size B should be a multiple of N.

pycolmap

Hi this is usually because the colmap is not installed correctly. You can notice that, in two logs, the colmap is loaded from different paths. Can you try re-install it?