Things to implement

[doloresgarcia]

Model features:

• Obtain number of hits and true particles per event distribution (2d histogram) and mean and variance [Gregor]
• Adding particle ID in the loss/model (see comments below) [Gregor; change to entire cluster pooling instead of the center node]
• check if the energy of the particle matches the energy of the hits that link to the particle and correct it [Gregor]
• train model gravnet, add track p as an input
• also check if the model would be best if the E pred was calculated from the sum of the particles inside the cluster.
• check if the loss function_E with the log works better (therefore without the 20 factor in the overall loss)
• implement inference by removing hits that have been already associated as in procedure 2 of the end to end paper

Plotting:

• visualize hits embedding space [Gregor]
• what hits is it picking as cluster center
• implement the plots, per type of particle (as in the paper) [Gregor]

[doloresgarcia]

Note: currently the clustering is implemented as x_alpha (so not doing the mean of the particles in the cluster) this could also be done but may be worth it to keep the simple approach for now

[doloresgarcia]

About adding particle ID: this can not be regressed directly because it should be obtained from the cluster.
Options:

• to pull the clusters and then regress 1 output for each subgraph
• to do some aggregation from per hit regression (this is more sloppy)

[doloresgarcia]

• higher q value [0.3] EASY
• sample different points in the class and add that to the loss Gregor
• Take out the batchnom/dropout. EASY NOW
• try other networks Dolo
• batchnorm before the output layer. NOW
• plot the gravnet clustering space Dolo
• Option to start with the physical space [NOT FOR NOW]