pytorch implementation for the paper
Mattias P. Heinrich, Max Blendowski, Ozan Oktay "TernaryNet: Faster Deep Model Inference without GPUs for Medical 3D Segmentation using Sparse and Binary Convolutions" currently under review for IJCARS MICCAI 2017 special issue
Currently, only the most basic training/validation example using ternary convolutions within a U-Net medical image segmentation pipeline are provided. This will be extended in the near future, also with the addition of Hamming distance optimised C-code for inference.
The proposed ternary hyperbolic tangent activation is defined as
m = torch.nn.Tanh() y = m((x*beta*2.0-beta))*0.5 y += -m((-x*beta*2.0-beta))*0.5
If you find the material useful please cite the above paper or contact me through my website mpheinrich.de