Training with GPUs

[raymondlo84]

Just tested the tensorflow-gpu, and you can actually run openvino and tensorflow-gpu side by side. Basically, you can use the GPU for training, and CPU for inference =)!

You have to install CUDA 10 (for Tensorflow 2.3 and 2.4) or CUDA 11 (for Tensorflow 2.5), and also the cudNN library. After that, you can modify the requirements.txt to change the line from tensorflow to tensorflow-gpu.
or you can do it manually

pip install --upgrade tensorflow-gpu
The key is you have to ensure the numpy==1.17.* is installed, or the model optimizer will fail...

so if anything you have to run this after the command above

pip install numpy==1.17.*

=) my 3090 RTX is pretty good =)

[raymondlo84]

Lastly, I have ported this to COLAB a while back for OAK-D camera, but the code is pretty much identical and you can use the COLAB GPUs as well... =)

https://colab.research.google.com/drive/1oNxfvx5jOfcmk1Nx0qavjLN8KtWcLRn6?usp=sharing