NVIDIA / MinkowskiEngine

Minkowski Engine is an auto-diff neural network library for high-dimensional sparse tensors
https://nvidia.github.io/MinkowskiEngine
Other
2.47k stars 367 forks source link

How to reuse kernel map between different networks #431

Open hellojxt opened 2 years ago

hellojxt commented 2 years ago

Hi, I read the documentation and know that the Minkowski Engine caches all kernel maps automatically. But I want to reuse the kernel map between different networks with the same input tensor to reduce overhead. Is this possible? or specifically, how to get the kernel map from a convolution (or transpose convolution) and set the kernel map to another convolution (or transpose convolution) with the same input shape?