NN clusterizer

• Invested time in training data selection and downsampling: NN now much better for regression
• Now also added momentum vector estimate. Shows good performance for pY/pX and pZ/pX
• Training for 2-class currently ongoing
• Efficiency and fake-rate ratios now much better and roughly reflect expectation

Momentum distribution, ideal:

--------------------------------------------------------------------

Momentum distribution, NN:

--------------------------------------------------------------------

Ratios:

GPU implementation

• Implemented cluster-splitting networks in clusterization code
• Pull requests: Will add the automatic detection of CUDA, MIGRAPHX, ROCM (currently on sepeparate branch)
• Build: Several changes to ONNXRuntime framework necessary to make build work. Seems to work now with CUDA and ROCM but non-matching version in al9_gpu container for CUB and THRUST cuda libraries...
  • Changed CMakeLists.txt not to through fatal if both builds are activated
  • Changed CreateGPUDataTransfer to seperate versions for CUDA and ROCM (CreateGPUDataTransferCUDA and CreateGPUDataTransferROCM)