Closed thucz closed 3 months ago
You should be able to switch the neural network components to FP16 using the steps here, since the code is based on PyTorch Lightning. diff-gaussian-rasterization
would be harder to convert to FP16 because you would have to change the CUDA code in a bunch of places and certain operations might not work with FP16 precision. I would start with the neural network parts of pixelSplat, since those are responsible for the vast majority of the memory usage.
In fact, I tried to train in precision=16. The diff-gaussian-rasterization
module will raise errors about the data type.
Hi! I want to train pixelsplat on higher resolution images. But the GPU memory limits my thoughts. Do you know how to adapt your pipeline especially
diff-gaussian-rasterization
module tofp16
training?