How to use --eval-mode for ns-train splatfacto for manually split the dataset

[Runasia]

I want to manually split the dataset into training and validation sets, and specify the selected images. To achieve this, I used the following command (a bit of it):

ns-process-data --data data/nerfstudio/xx/train && --eval-data data/nerfstudio/xx/test

Then, I ran the training with:

ns-train splatfacto --data data/nerfstudio/xx --eval-mode filename --max-num-iterations 200001 --steps-per-save 10000 --save-only-latest-checkpoint False --steps-per-eval-all-images 10000

than i find another command

ns-train splatfacto --data data/nerfstudio/xx--max-num-iterations 200001 --steps-per-save 10000 --save-only-latest-checkpoint False --steps-per-eval-all-images 10000 --pipeline.datamanager.eval-image-indices 0 1 --pipeline.datamanager.data data/nerfstudio/xx

But when I use ns-eval to generate rendered images, it gives me three images, which is beyond my control. （just like i set --pipeline.datamanager.eval-image-indices 0 1 ）

Because I copied the data state to get COLMAP (when the image count is low, I couldn't proceed, so I took a few images for validation and copied other images for training), I want to evaluate my data.

[basebala]

I have addressed this issue in this PR: #3668

What I found upon an inspection of the codebase, is that eval-image-indices is declared as part of the datamanager config but not used anywhere. I'm also unsure how you used that parameter with splatfacto, because splatfacto does not use the Vanilla Data Manager Config (nerfacto does, however). My PR allows runs of either of these models and more (any using NerfstudioDataParserConfig) to manually input a list of indices for evaluation.

[Runasia]

I have addressed this issue in this PR: #3668

What I found upon an inspection of the codebase, is that eval-image-indices is declared as part of the datamanager config but not used anywhere. I'm also unsure how you used that parameter with splatfacto, because splatfacto does not use the Vanilla Data Manager Config (nerfacto does, however). My PR allows runs of either of these models and more (any using NerfstudioDataParserConfig) to manually input a list of indices for evaluation.

Thank you for the guidance! I'll pull the changes and test again. Really appreciate your help