Performance of NN. Data pipeline and input structure.
- TensorFlow with Keras
- Training and testing based on simulation for: CERN 18
- run 403, 7 GeV/c pions @ 20-140 degree polar angle
Data pipeline:
![](/~rdzhigad/data/prt/nn_performance/nn_pipeline.png)
Index based data translated into full dimensional tensors. Example of one event @ 20 degree polar angle. X dimension has size of 532 (512 for channel ids and 20 to encode momentum/angle id)
![](/~rdzhigad/data/prt/nn_performance/tp_19.5.png)
Performance of 1-layer dense NN in comparison to Time Imaging. Each angle was trained with separate NN:
![](/~rdzhigad/data/prt/nn_performance/draw_ai_comp/comp_sep_0.5.png)
Efficiency:
![](/~rdzhigad/data/prt/nn_performance/draw_ai_comp/comp_eff_0.5.png)
One common NN trained on 20 and 140 degree data:
![](/~rdzhigad/data/prt/nn_performance/draw_ai_comp/comp_sep_20_140.png)
One common NN trained on 20,30,40 ... 140 degree data (10 degree step):
![](/~rdzhigad/data/prt/nn_performance/draw_ai_comp/comp_sep_10.png)