internal/dataset.py

[Miner-AI]

internal/dataset.py

[SuLvXiangXin]

@Miner-AI If you want to train your model with images which are downsampled by 4, you need to have images_4 exists. You can checkout https://github.com/SuLvXiangXin/multinerf-pytorch/blob/main/scripts/local_colmap_and_resize.sh for preparing your own dataset. Or you can simply specify --gin_bindings="Config.factor = 0" to run without downsampling.

[SuLvXiangXin]

By the way, I have implemented another powerful method called "Zipnerf", which is available at https://github.com/SuLvXiangXin/zipnerf-pytorch

[Miner-AI]

Thank u for reply！
I have another question about internal/coord.py. When I try to use llff_raw.gin to train on rawnerf's data, I met an error about the method construct_ray_warps(internal/coord.py), it is caused by missing definition of Model.raydist_fn in configs/llff_raw.gin(which in 360.gin, Model.raydist_fn = @torch.reciprocal). How can I solve this problem？

appreciate your reply！thanks a lot！

[SuLvXiangXin]

I tested rawnerf can run through. Because this repo simply translates the official jax version into pytorch for implementing zipnerf, it has not studied the principle of rawnerf in depth. It is okay to not define model.raydist_fn in llff_raw.gin, which means that sampling on rays is simply equidistant sampling instead of reciprocal. If you are interested in this project, you can also pull request to improve it. If you encounter any other errors
, please post the error message in the issue