KataGo
KataGo copied to clipboard
lightvector
Upgrade to tensorflow 2.4.0
So that could use CUDA 11.0 + tensorflow_gpu 2.4.0, makes synchronous training on a single machine possible.
Does this change upgrade all python scripts to run with tensorflow 2.4?
I also need to upgrade python scripts to run with tensorflow 2.9 because it is the only version that can be installed in my Macbook M1 Pro.
If all python scripts are upgraded to tensorflow 2.4 or 2.9, it gives me a chance to see if we can utilize Apple neural engines to speed up katago by converting a tensorflow network to Apple's Core ML model in a Macbook M1 Pro.
That's OK.
I found a method to convert a TensorFlow 1.x network to Apple's Core ML model on Google Colab, so I don't have a strong requirement to upgrade to TensorFlow 2.x now.
That's OK.
I found a method to convert a TensorFlow 1.x network to Apple's Core ML model on Google Colab, so I don't have a strong requirement to upgrade to TensorFlow 2.x now.
How did you do that? Can you share the model?
That's OK. I found a method to convert a TensorFlow 1.x network to Apple's Core ML model on Google Colab, so I don't have a strong requirement to upgrade to TensorFlow 2.x now.
How did you do that? Can you share the model?
See my recent release page: https://github.com/ChinChangYang/KataGo/releases/tag/v1.11.0-coreml3 For more information, see this old release page: https://github.com/ChinChangYang/KataGo/releases/tag/v1.11.0-coreml2
@ChinChangYang Thank you! BTW, What is a Neural Engine Core? Apple says M1-series has 16 Neural Engine Core. However, we can only set two in Katago?
@ChinChangYang Thank you!
BTW, What is a Neural Engine Core? Apple says M1-series has 16 Neural Engine Core. However, we can only set two in Katago?
The M1-series has 16 Neural Engine cores, but you can only set two threads in KataGo because it's a limitation of the code. However, you can modify the code to run more threads for the Neural Engine.
The two threads in KataGo run two identical Core ML models, and then Core ML decides which compute units to use for each operation in the neural network. These compute units can be the CPU, GPU, or Neural Engine (NE). In my experience, KataGo's Core ML only selects the CPU or NE, and I've noticed that the NE is kept busy by the two threads running Core ML. It's up to you whether you want to try using more threads to see if it improves performance.