This derivation can be used in conjunction with pytorch-bin to avoid the long build time of PyTorch/TorchVision CUDA builds.