Open Tasmay-Tibrewal opened 9 months ago
Hi sorry I'm late to the party but I'm going to leave this comment here for anyone that's facing memory problems with complex datasets (such as the multi-view Plenoptic/DyNeRF data set),
I've also encountered it with pretty much every other dynamic NeRF. Even if you load the training data on the GPU there's not enough memory. I suggest you modify the code to actively load the data during training, though this will significantly decrease the speed of the method (as you'll be loading the training data from storage memory rather than RAM or GPU memory). The other option is purchasing more RAM or using a compute system such as AWS, Kaggle, etc.
We are trying to reproduce the results on the plenoptic video dataset, but it is only using the cpu and not the gpu.