New developments:

• Added local occupancy estimator (this was one request by GSI in the group meeting):
  vector<float> loc_occ(all(pad_time_bins), 0);
  for every pad_time_bin in tpc_sector{
  for pad_time_bin_2 in [time(pad_time_bin) - 20, time(pad_time_bin) + 20] and all_pads_in_ROC(pad_time_bin){
  loc_occ[pad_time_bin] += (int)(pad_time_bin_2 > 0);
  }
  loc_occ[pad_time_bin] /= (all_pads_in_ROC(pad_time_bin) * 41);
  }
• Distortion simulation working now: full-system-test, distortion type 2 + slow-realisitic-full-sim + runnumber=311000
• Got ITS-TPC matching + QC to work using the reco_NOGPU.log (WORKFLOWMODE="print")
• Simulated different IR's and collision systems now (PbPb: 50kHz, 30kHz, 15kHz; pp: 1500kHz, 1000kHz, 500kHz)
• Created dedicated dataset with loopers: 1 Ev. Pb-Pb 500kHz +  lowneut=true = many loopers at high time-bins 
• Created dedicated simulation for overlap training data: 5 Ev. Pb-Pb 200 kHz should give a high local occupancy and a lot of overlapping tracks (to be seen if it is locally realisitic)

Performed investigation on spacecharge distorted data.

• Tested NN's trained on not-spacecharge-distorted data to see if they are good enough or need retraining.
• Tested different input sizes and (NN classifciation + NN regression) vs. (NN classification + GPU CF regression) vs. (pure GPU CF regression)
• Overall results:
  • NN classification + NN regression:
    • Regression does not work (well), bands are significantly broader or look like noise. Needs retraining. Was basically expected.
    • When applying all corrections, charge offset of native clusterizer does not appear anymore (investigating)
  • NN classification + GPU CF:
    • Similar effects as in non-distorted data: ~10-20% reduction of clusters. Bands look virtually identical for all variables (CoG pad and time, sigma, etc.) between both algorithms -> So classification seems to be somewhat robust to SC distortion

Plan ahead

• Retrain NN's with SC-distorted data and perform QA again
• Streamline everything more. Currently, many manual steps -> focus on automatization