The GraphTemplate Source Files

Similar to the event_template class, the graph_template class allows for graphs to be constructed, that can be later used to train a Graph Neural Network. In this class, specific types of particles can be used to construct Nodes, which are subsequently assigned some set of input features. Assignment of features is performed by passing lambda functions that specifically retrieve such features. For instance, given some callable function , with input particle and output feature y_i, the template class will assign each particle node the value in a graph. By a similar token, to construct edge features, the callable function can be functionally expressed as .

To further simplify the pipeline, internal functions exploit dimensional analysis on the output vector space to infer the dimension of the PyTorch tensor and simply builds this tensor.

A simple code implementation would look structurally as below:

#ifndef GRAPHS_GRAPHNAME_H
#define GRAPHS_GRAPHNAME_H

#include <event-name/some-event.h>

class graph_name: public graph_template
{
     public:
         graph_name();
         ~graph_name();

         graph_template* clone() override; // make the implemenation cloneable
         void CompileEvent() override; // Compiler space.

};

#endif

#include "graph-name.h"

// boiler plate code
graph_name::graph_name(){this -> name = "graph-name";}
graph_name::~graph_name(){}
graph_template::clone(){return (graph_template*)new graph_name();}

void graph_name::CompileEvent(){
     auto some_graph_fx = [](bool* out, event_name* ev){*out = ev -> truth_signal_variable;}; // is the event a signal event?

     auto some_node_fx = [](int* out, particle_template* p){*out = p -> is_top;}; // is this particle for example a top?

     auto some_edge_fx = [](int* out, std::tuple<particle_template*, particle_template*>* e_ij){
         *out = std::get<0>(*e_ij) -> top_index == std::get<1>(*e_ij) -> top_index; // define truth edge topology
     };

     auto some_other_graph_fx = [](double* out, event_name* ev){*out = ev -> missing_et;}; // get missing transverse momenta

     auto some_other_node_fx = [](double* out, particle_template* p){*out = p -> pt;}; // get particle pt

     auto some_other_edge_fx = [](double* out, std::tuple<particle_template*, particle_template*>* e_ij){
         *out = std::get<0>(*e_ij) -> pt - std::get<1>(*e_ij) -> pt; // pt difference between particles
     };


     // get the event that has all the fill event data
     event_name* event = this -> get_event<event_name>();
     std::vector<particle_template*> particles = event -> some_var_with_particles;

     // ------ define a prior topology or leave this out for a fully connected graph ----- //
     // for instance, if the target particle is a top, then no two b-jets can share the same top.
     auto bias_topology = [](particle_template* p1, particle_template* p2){return p1 -> is_b != p2 -> is_b;};
     this -> define_topology(bias_topology);  // this topology bias will be reflected in subsequent topology functions.

     // ----- truth features ----- //
     this -> add_graph_truth_feature<bool, event_name>(event, some_graph_fx, "name-this-truth");
     this -> add_node_truth_feature<int, particle_template>(some_node_fx, "is_top");
     this -> add_edge_truth_feature<int, particle_template>(some_edge_fx, "top_edge");

     // ----- observables --------- //
     this -> add_graph_data_feature<double, event_name>(event, some_other_graph_fx, "some-other-data");
     this -> add_graph_data_feature<double, particle_template>(some_other_node_fx, "pt");
     this -> add_graph_data_feature<double, particle_template>(some_other_edge_fx, "delta-pt");

     // The class also allows for access to event data that relates to the origin root file.
     std::string root_file_path = this -> filename;

     // or event_t struct access
     event_t ev_data = this -> data;
}

The above code is all that is required to construct graphs for Graph Neural Network training! There are no additional steps required to start using the PyTorch API.

For more information about the methods and attributes of the graph_template class, see the core-class documentation GraphTemplate Methods.