The ModelTemplate Source Files

To add a new model to the framework, navigate to the template files (src/AnalysisG/templates/model) and copy the source files into the existing model directory. Make sure to rename the folder appropriately, and rename files with the <model-name> prefix as needed. The given template files have all the needed structures in place, and essentially just require search and replace modifications.

The C++ Source Code

#ifndef <model-name>_H
#define <model-name>_H
#include <templates/model_template.h>

class <model-name>: public model_template
{
    public:
        <model-name>();
        ~<model-name>();
        model_template* clone() override;
        void forward(graph_t*) override;

        torch::nn::Sequential* example = nullptr;
};

#endif

#include <model.h>

model::model(){

    this -> example = new torch::nn::Sequential({
            {"L1", torch::nn::Linear(2, 2)},
            {"RELU", torch::nn::ReLU()},
            {"L2", torch::nn::Linear(2, 2)}
    });

    this -> register_module(this -> example);
}

void model::forward(graph_t* data){

    // fetch the input data of the model.
    // If the variable is not available, this will return a nullptr.
    torch::Tensor graph = data -> get_data_graph("graph") -> clone();
    torch::Tensor node  = data -> get_data_node("node") -> clone();
    torch::Tensor edge  = data -> get_data_edge("edge") -> clone();
    torch::Tensor edge_index = data -> edge_index -> clone();

    // output the prediction weights for edges, nodes, graphs.
    this -> prediction_graph_feature("..."; <some-tensor>);
    this -> prediction_node_feature("...", <some-tensor>);
    this -> prediction_edge_feature("...", <some-tensor>);
    if (!this -> inference_mode){return;} // skips any variables not avaliable during inference time.
    this -> prediction_extra("...", <some-tensor>);  // Any variables that should be dumped during the inference.
}

model::~model(){}
model_template* model::clone(){
    return new model();
}

Cython Interface Files

The code below would be the interface of the model via Cython which can be initialized from the Python interpreter Similar to C++, they require a header (.pxd) and source file (.pyx).

class model.pxd

# distutils: language=c++ # cython: language_level=3

from libcpp cimport int, bool from AnalysisG.core.model_template cimport model_template, ModelTemplate

cdef extern from “<models/model.h>”:

cdef cppclass model(model_template):

model() except+

cdef class ExampleModel(ModelTemplate): pass

class model.pyx

# distutils: language=c++ # cython: language_level=3

from AnalysisG.core.model_template cimport ModelTemplate from AnalysisG.models.<model-name> cimport ExampleModel

cdef class ExampleModel(ModelTemplate):

def __cinit__(self): self.nn_ptr = new model() def __init__(self): pass def __dealloc__(self): del self.nn_ptr