Tensorflow.NET on ARM64

[ianier]

I opened this issue in the TF repository to allow Windows ARM64 as a compilation target (on an x64 host, to be clear), and I was wondering if there were similar ongoing initiatives from the Tensorflow.NET community. If not, are you guys interested in helping with it?

[minhduc66532]

This guy already built the Tensorflow 2.3.1 C++ API on the Nvidia Jetson Nano (using Jetpack 4.4). You can just scroll down until u found this: Download the package, unzip it, rename the libtensorflow.so.2.3.1 to libtensorflow.so. Drag it into your app, run it on the terminal ./YourProgramName. Also in the project file, change the RuntimeIdentifier to linux-arm

<PropertyGroup>
    <OutputType>Exe</OutputType>
    <TargetFramework>net5.0</TargetFramework>
    <RuntimeIdentifier>linux-arm64</RuntimeIdentifier> <---
    <SelfContained>true</SelfContained>
    <PublishSingleFile>true</PublishSingleFile>
    <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
  </PropertyGroup>

IF u DO NOT use the Jetson Nano (which has a Nvidia GPU on it) then u don' need CUDA and CuDNN. But that is my guess since i haven't tested it on a raspberry pi. If u want to i can do it for ya System informations (You need the exact CUDA version in order to run this, CuDNN ? Haven't tested it yet):

• Ubuntu 18.04
• CUDA 10.2
• CuDNN 8.0

[ianier]

@minhduc66532: Please note that this issue is about compiling TF for full Windows on ARM (e.g. the Surface Pro X), which would be an analogous effort to what they're doing to get TF to compile for the Apple M1 CPU.

[minhduc66532]

Hmm that suck i misunderstood the question lol, sorry about that :(