Recursive Graph Neural Network (`models.RecursiveGraphNeuralNetwork`)

Import with:

from AnalysisG.models.RecursiveGraphNeuralNetwork import RecursiveGraphNeuralNetwork

RecursiveGraphNeuralNetwork

RecursiveGraphNeuralNetwork is a ModelTemplate subclass wrapping the recursivegraphneuralnetwork C++ class.

The model performs recursive message-passing over node and edge features using 8 sub-networks to simultaneously predict edge-level top assignments, node-level aggregated features, and event-level exotic resonance and number-of-tops outputs.

Architecture (from `RecursiveGraphNeuralNetwork.cxx`)

Constructor signature: recursivegraphneuralnetwork(int rep=1024, double drop_out=0.1)

Sub-network	Role
`rnn_dx`	Edge message network: `(_dx + 2×_rep) → (_rep×2) → _rep`; LayerNorm + SiLU + Dropout + Sigmoid
`rnn_x`	Node encoder: `(_x + _rep) → (_rep×2) → _rep`; LayerNorm + Dropout
`rnn_merge`	Hidden-state merge: `(_rep×3) → _rep`; LayerNorm + SiLU + Dropout
`rnn_update`	Edge-prediction updater: `(_output×2 + _rep×2) → _output`; LayerNorm + SiLU + Dropout
`exotic_mlp`	Exotic resonance head: `(_x + _output) → (_rep×2) → _output`; LayerNorm + Dropout
`node_aggr_mlp`	Node aggregation: `_x → _rep → _x`; LayerNorm + Dropout
`ntops_mlp`	Number-of-tops head: `(_x + 4) → _rep → _x`; LayerNorm + SiLU + ReLU
`exo_mlp`	Second exotic head (post-aggregation): same pattern as `ntops_mlp`

Architecture dimensions:

Field	Default	Meaning
`_dx`	26	Input edge-feature dimension (concatenated node features for both endpoints)
`_x`	5	Input node-feature dimension
`_output`	2	Output edge-prediction dimension (binary: same-top or not)
`_rep`	256	Internal hidden-state dimension (overridden by constructor argument)

Hyper-parameters (Python layer)

Property	Type	Description
`drop_out`	`float`	Dropout probability. Default `0.1`.
`res_mass`	`float`	Target resonance mass [MeV] used as an auxiliary loss constraint. Default `0.0` (disabled).
`is_mc`	`bool`	Include MC-truth features in the input. Default `True`.

Usage

from AnalysisG import Analysis
from AnalysisG.models.RecursiveGraphNeuralNetwork import RecursiveGraphNeuralNetwork

ana = Analysis()
ana.AddModel(RecursiveGraphNeuralNetwork)
ana.Epochs = 20
ana.Start()

C++ Reference

class recursivegraphneuralnetwork : public model_template

Public Functions

recursivegraphneuralnetwork(int rep = 1024, double dpt = 0.1)

~recursivegraphneuralnetwork()

virtual model_template *clone() override

Create a default-constructed clone. Override in subclasses.

Returns:: New model instance.

virtual void forward(graph_t*) override

Execute one forward pass. Override in subclasses to implement the model computation.

Parameters:: data – Pointer to the input graph.

torch::Tensor message(torch::Tensor trk_i, torch::Tensor trk_j, torch::Tensor pmc, torch::Tensor pmc_i, torch::Tensor pmc_j, torch::Tensor hx_i, torch::Tensor hx_j)

Public Members

int _dx = 26

int _x = 5

int _output = 2

int _rep = 256

double res_mass = 0

double drop_out = 0.1

bool is_mc = true

torch::nn::Sequential *rnn_x = nullptr

torch::nn::Sequential *rnn_dx = nullptr

torch::nn::Sequential *rnn_merge = nullptr

torch::nn::Sequential *rnn_update = nullptr

torch::nn::Sequential *exotic_mlp = nullptr

torch::nn::Sequential *node_aggr_mlp = nullptr

torch::nn::Sequential *ntops_mlp = nullptr

torch::nn::Sequential *exo_mlp = nullptr

Recursive Graph Neural Network (models.RecursiveGraphNeuralNetwork)