jc211
Saving a Model Snapshot
Hi, I would like to save a snapshot of a model trained with ParticleNeRF. I tried to use testbed.save_snapshot method. However, if I save a model like this and then reload it into the instant-ngp app with --load_snapshot I see only a ghost-like shape of an object
- example for the wheel dataset (the ghost-like shape is barely visible):
From what I have checked it does not matter if the scene is dynamic or not - I get similar behaviour on the fox dataset.
If I restart the training from this ghost-like state the model usually converges more quickly, so clearly some information is saved.
Is there a way to save a ParticleNeRF model in such a way that one has a fully ready visualisation after reloading a snapshot - without the need to restart the training?
Unfortunately, the save snapshot was not implemented - mainly because it was not needed for the original paper. The network however is being saved which explains the ghosting.
@jc211 are you planning to implement this in the future?
If not, could you point me on how one could go about implementing this?
I don't have any plans to at the moment. I can help guide the process if you want to have a go.