CNN with only axis X and Y¶

Conclusion: no convergence with only the X and Y. the addition of the Z axis is essential

CNN training with 3 axis¶

Conclusion:

• Convergence is struggling to begin, 30 iterations without progress

• from iteration 40 the convergence is very slow

• 150 itérations take 10mn with 4 CPU on my laptop

• try to add iteration at CCIN2P3 with GPU

CNN test¶

Create files with probabilty¶

• bkg_dataset_nutrig_gp13_test_seed_300.npz
• sig_dataset_nutrig_gp13_test_seed_300.npz

with same index order

Time for inference¶

Inference CPU time= 1.52 s, for 10000 traces on laptop 4 CPUs

Plot distribution of CNN proba for signal/background¶

Separability index¶

index_sep= 0.50

To compare with the ICRC dataset: 0.87

with définition

https://github.com/grand-mother/NUTRIG1/blob/7efde4e45f8d00259e0f93e66022efc00a2a24a3/src/nutrig/eval/cnn_ref.py#L136

Fix bug separability index: now surface intersection is really computed

Plot separability index versus SNR threshold¶