Memorandum: When I tried TensorFlow with Tribuo, it didn't work, so I went on a journey to find the head family and lost.

Before

2020/10/03 This is an impression at the time of writing.

The story of TensorFlow in Tribuo doesn't come up here. Not only Tribuo, but various things are tough. It's just an addictive record. This is a problem before linking TensorFlow from Tribuo.

Conclusion

TensorFlow Java must also be built with Java 7 in the latest version of the Maven Central Repository (1.15.0). I don't want to do a lot of new projects that adopt Java 7. It worked until the build with Java11, but with Java12 or later, an error occurred at runtime.

Even the latest version of TensorFlow (2.3.0) did not support Cuda11. If you want to move it forcibly, you can move it.

I forgot Tribuo. It was morning when I noticed it, and before I knew it, I forgot about Java in TensorFlow and implemented python ... After all, I don't like scripting languages.

It is possible to use GPU from TensorFlow on Windos, but Windows on WSL2 on ubuntu did not work in my environment.

Understanding the current situation

Docer Desktop for Windows does not support GPGPU

I put Linux in Docer in a Windows environment and tried to use GPU from TensorFlow, but it didn't work. In the first place, the Windows version of Docer Desktop did not support GPGPU.

I ran Docker on Ubuntu built on WSL2 on Windows, but the GPU doesn't recognize it

I put Docer in Ubuntu built on WSL2 on Windows and tried it because there was information like Haha, but it was not good in my environment. In the first place, even on Ubuntu for installing Docker, if you execute / usr / local / cuda / samples / 1_Utilities / deviceQuery / deviceQuery, 35 will be returned.

I also tried the TensorFlow image, but it didn't work either.

GPU does not recognize on Ubuntu built on WSL2 on Windows

AMD:Ryzen 3700x GPU:RTX3090 OS:Windows Build 20226.rs_prerelease.200925-1415 NVIDIA GPU Computing Toolkit:CUDA11.1/CUDA10.1/CUDA10.0 CUDANN:CUDA11.1/CUDA10.1/CUDA10.0

I was able to use the GPU from Ubuntu built on WSL2 on Windows, and there was information that it actually worked, so I tried it, but all the toolkits of CUDA10 / CUDA11-0 / CUDA11-1 are my It didn't work in the environment. ~~ Because Gefoce RTX 3090? AMD's curse? ~~

When you hit DeviceQuery, error code 35 is output.

Built directly on the host OS Windows 10

If you search for this, information will come out, so I was able to build it without much trouble. I don't want to build a development environment on a host OS that is not a virtual environment because I create a virtual environment for each required environment. I wish I could use the GPU with Hyper-v on windows, but the host OS is not Windows Server, WSL2 didn't work, and only time passed.

TensorFlow Java also worked on the GPU.

However, if Java 12 or later is specified, an error will occur at runtime.

Impressions

The version relationship is chaos.