Training beyond 18k iterations throws lot of errors

[Jaykumaran]

Hi,

Any particular reason that the most part of code is assumes that the training would be 18k iters and something similar for mesh updates as well? As i see the training plots saved with logging i can oobserve that the mesh is definitely improving in later part of 12k. So why there isnt any flexible option to make use of the arg I'm passing in cli instead it always uses 18k iters from configs/fast

To note that i tried to modify few internals of the code and yaml files to adapt it for 30k iters either the loss values went NAN or OOM error.

One possible intution that could be as mentioned in other mesh rendering papers where they make use of 7k iter vanilla model instead of 30k iters because the gaussian would have been almost thin trying to increase SSIM at 30k iters whereas a 7k checkpoint is a sweet spot as the gaussian would have estimated the reconstruction priors of the scene but not yet fully optimized for SSIM , would be of a size thats suitable for Mesh rendering -- I think it was from SuGAR paper, not sure though..