Grift Model (`models.grift`)

The Grift Graph Neural Network is a message-passing GNN for top-quark combinatorics. Import with:

from AnalysisG.models.grift import Grift

Grift

Grift is a ModelTemplate subclass wrapping the grift C++ class.

Architecture (from `grift.cxx`)

Grift uses 7 torch.nn.Sequential sub-networks arranged in a recurrent message-passing loop:

Sub-network	Role
`rnn_x`	Node encoder: `(_xin + _xrec) → (_xrec + _xin) → _xrec`; uses LayerNorm + Tanh + LeakyReLU
`rnn_dx`	Edge message network: `(_xrec × 3) → (_xrec × 2) → _xrec`; Tanh activations
`rnn_hxx`	Hidden-state aggregator: `(_xrec×2 + 2) → (_xrec×2) → _xrec`; LeakyReLU + Tanh
`rnn_txx`	Top-edge predictor: `(_xrec × 3) → (_xrec × 2) → _xout`; LeakyReLU + Tanh
`rnn_rxx`	Resonance predictor: `(_xrec×4) → _hidden → _xout`; LeakyReLU + Sigmoid
`mlp_ntop`	Number-of-tops readout: `(_xtop + _xrec) → _xrec → _xtop`; LayerNorm + LeakyReLU + Sigmoid
`mlp_sig`	Signal classifier: `(_xout×2 + _xrec×4 + _xtop×2) → (_xrec×2) → _xout`; LayerNorm + LeakyReLU + Sigmoid

Architecture dimensions (C++ defaults):

Field	Default	Meaning
`_hidden`	1024	Width of the `rnn_rxx` hidden layer
`_xrec`	128	Recurrent hidden-state dimension used by all sub-networks
`_xin`	6	Input node-feature dimension (`pt, eta, phi, E, is_lep, is_b`)
`_xout`	2	Output edge/graph prediction dimension (binary classification)
`_xtop`	5	Number-of-tops feature dimension

Hyper-parameters (Python layer)

Property	Type	Description
`drop_out`	`float`	Dropout probability applied to hidden layers. Default `0.0`.
`is_mc`	`bool`	When `True` MC-truth features are included in the input tensor. Default `True`.
`pagerank`	`bool`	When `True` a PageRank score is computed for each node during inference and stored in the output node tensor. Default `False`.

Expected Input Features

Grift reads the following tensors from graph_t inside forward():

Feature	Slot	Registered by
`"pt"`	node data	graph class
`"eta"`	node data	graph class
`"phi"`	node data	graph class
`"energy"`	node data	graph class
`"is_lep"`	node data	graph class
`"is_b"`	node data	graph class
edge index	COO	graph class

Usage

from AnalysisG import Analysis
from AnalysisG.models.grift import Grift

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

C++ Reference

class grift : public model_template

Public Functions

grift()

~grift()

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 hx_i, torch::Tensor hx_j)

torch::Tensor node_encode(torch::Tensor pmc, torch::Tensor num_node, torch::Tensor *node_rnn)

torch::Tensor recurse(torch::Tensor *node_i, torch::Tensor *idx_mat, torch::Tensor *edge_index_, torch::Tensor *edge_index, torch::Tensor *edge_rnn, torch::Tensor *node_dnn, torch::Tensor *top_edge, torch::Tensor *pmc, torch::Tensor *node_s)

Public Members

int _hidden = 1024

int _xrec = 128

int _xin = 6

int _xout = 2

int _xtop = 5

bool is_mc = true

bool init = false

bool pagerank = false

double drop_out = 0

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

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

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

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

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

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

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

torch::Tensor x_nulls

torch::Tensor dx_nulls

torch::Tensor te_nulls

Grift Model (models.grift)