In this experimental release with prebuilt binaries it’s about 5Mb per architecture. This includes all operators for inference (that is forward only). We’re working on selective compilation so that you can build a smaller bundle with only a subset of ops that you use. With that for common CNNs it should get to 1-2 Mb range or even smaller.
Comments
In this experimental release with prebuilt binaries it’s about 5Mb per architecture. This includes all operators for inference (that is forward only). We’re working on selective compilation so that you can build a smaller bundle with only a subset of ops that you use. With that for common CNNs it should get to 1-2 Mb range or even smaller.