Mask_scales miss few pt files that fails train_contrastive_feature.py training.

[XcloudFance]

Hi teams, when I tried deploying this project in my machine, downloading pretrained 3dgs models into local, few parameters are missing to be set and I did them according to the code.

However, after preparing data, I tried to run train_contrastive_feature.py to train. The results shows : No such file or directory : "xxx/mask_scales/DSCxxx.pt' where it randomly appears. So I checked the folder and indeed it wasnt there.

The dataset I am using is Mips-NeRF-360. I also tried two scenes: garden and bonsai. Both of them miss some pt files when generating mask_scales.

Any idea how to fix it?

[Jumpat]

Hi, sorry for the late response. You need to extract these mask scales by yourself. Please check the prepare data section in our README.

[widifadi]

Hi I'm currently also facing the same problem. I used my own dataset, everything is fine until creating SAM mask. My input has 147 images so does the SAM mask output, but after running get_scale.py the output *.pt files are only 128 files. Therefore running train_contrastive_feature.py showing FileNotFoundError and I cannot continue.

I'm also not sure if this is also the cause, I've moved my model outputs to another drive due to space issue, but I didn't change the structure, and therefore the cfg_args model_input param is different to the current model_input path. But I think it should not be a problem since I specified the model path when running get_scale.py:

python get_scale.py \
> --image_root /workspace/data/<my data> \
> --model_path /workspace/render/<my model>

Looking for config file in /workspace/render/<my model>/cfg_args
Config file found: /workspace/render/<my model>/cfg_args
Loading trained model at iteration 30000, None
Allow Camera Principle Point Shift: False
Reading camera 147/147
Loading Training Cameras
Loading Test Cameras
147it [01:17,  1.90it/s]
0it [00:00, ?it/s]/workspace/SegAnyGAussians/gaussian_env/lib/python3.8/site-packages/torch/functional.py:478: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at  ../aten/src/ATen/native/TensorShape.cpp:2894.)
  return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]
128it [00:47,  2.70it/s]

I'm concerned with the 'None' when loading the model at the specified iteration, is it because I'm moving my model?

[widifadi]

I found the solution. Apparently, if you train using --eval flag, the get_scale.py will only scale the training input image instead of the whole input.