Failed to run train_contrastive_feature.py

[L1n2hang]

Thanks for your excellent work on implementing SegAnything on GS dataset. But I did have some issue while running the code. Here's the error I had: """ SegAnyGAussians\train_contrastive_feature.py", line 75, in training feature_gaussians = FeatureGaussianModel(dataset.feature_dim) AttributeError: 'GroupParams' object has no attribute 'feature_dim' """ I would greatly appreciate for your suggestions :)

[L1n2hang]

emm... seems like it's working? Here's my "solution":

[Jumpat]

😂Hi, I wonder how it works in fact. Just add --feature_dim 32 should make sense. Dont set allow priciple point shift as True. This may make the learned feature have some bias.

[L1n2hang]

Oops, thanks for sharing that. I might rerun the training process after this one to see if there are any differences in the learned features.

A few quick follow-up questions:

1. What would be an average training duration? Mine seems to take about 44 hours to complete the entire training process. Could this be because the allow_principle_pt = True setting might add extra hours?

2. The configuration from my dataset for Gaussian splatting doesn't include an attribute named feature_dim. Is this because I'm using my own dataset, or is it missing from another configuration? I would greatly appreciate it if you could provide a sample configuration so I can compare it with my own.

Thanks :)

[Jumpat]

Hi, sorry for the late response. 44 hours is far beyond my expectation. On my device (single 3090) the training can be finished within one hour. Allowing principle point shift does not affect the training time.

The feature_dim attribute is a new attribute added by SAGA, which makes our config slightly different from the original GS. This has nothing to do with the dataset.