I am using AWC inf1.xlarge EC2 machine and doing experiments with YOLOv4 object detection model. I tried to run this example that uses TensorFlow-neuron 1.15, everything worked without any difficulties, achieved pretty similar results as given in the tutorial. Unfortunately, TensorFlow 1.15 is deprecated and this version is not getting any fixes, updates and etc.
I saw that there is TensorFlow 2 for neuron chips, so thought that there should be any difficulties getting similar results. So, not to invent the wheel or not create a pretty different YOLOv4 implementation as used in the mentioned example, I ported the TF1.15 YOLOv4 model to TF2, this was pretty easy.
So, the next step was to compile it for neuron, and here problems started, it seems the compiler doesn't give any feedback if it compiled it correctly or not, takes a lot of time, sometimes even EC2 machine hangs and I need to do force stop -> start, to get it working again.
Then I started to clip parts of YOLOv4, thinking that maybe there are unsupported operations or etc. But still, it compiles but never runs on all cores, simply on one core...
So, my question is, maybe this TF2 neuron package is missing some kind of documentation, how to correctly compile the model, or how to debug it? It's really annoying that I can't get feedback about the compilation process. Does it mean that to use these neuron chips is better move back in time to TF1.15?