Extract track parameters for each cluster and check performance e.g. pT differential
1. Neural network performance in 3D
Testing 7x7x7 input, boundary implemented between IROC and OROC1 (due to different pad sizes)
Comparison for classification with 2D case with variation of threshold
3D network has clear potential to distinguish between fake and real clusters
Fake-rate falls steeper than efficiency -> Slight compromise in efficiency can lead to strong reduction in fake-rate
2. Check neural network speed on GPU
Goal for Run 4: 50 mio. clusters / s of processing speed
With improvements for GPU's and parallel processing, goal to be reached now ~10-20 mio. clusters / s
Neural network classification: Float16 should be good enough, training scripts implemented & working
With some model optimizations (e.g. matrix multiplies as multiples of 4 or 8 for tensor optimizations): Reaching ~28 mio. clusters / s (fp16) and ~22 mio. clusters / s (fp32) on Nvidia Tesla V100 PCle (aliceml server)
3. Implementation of looper tagger
Create vector of size (max(time) / granularity, row, pad)
Accumulate all MC labels within "granularity" into this vector
Check for t in time_window: for p in pad_window: if any label occurs more than n times: tag region charge_map[row][pad][time : time + time_window] as looper
Need to add check for charge decrease -> low momentum particles can have maxima in ion tail
