The Model Class

A module which wraps the given model into a Cython compatible object. Following the mapping procedure, the model is given additional attributes and functionality, which includes some minor adjustments to PyTorch Geometric.

AnalysisG.Model.Model(model = None):
Parameters:

model – A trainable model.

match_data_model_vars(Data sample) → None

Inspects the input sample and creates an internal map where the inputs and outputs of the model are mapped to the variables of the sample.

backward()

Triggers the back-propagation of the learnable parameters.

save()

Saves the current state of the model’s trainable parameters.

load()

Restores a saved training state.

__call__(data) → dict

Executes the model on the input data and outputs a dictionary with the following keys:

  • L_*: The loss associated with the particular output of the model.

  • A_*: The accuracy associated with the model’s prediction (if it is of classification type).

  • total: The total loss of the all the model outputs \mathcal{L}_{total} = \Sigma_{i} \mathcal{L} _i

  • graphs: Unpacks a batched data element with the output of the model’s output. The unpacking is a hacked method of the to_data_list() method of PyGeometric.

Variables:

  • __param__ (dict) – A free parameter dictionary used to initialize the model with. This parameter relevant if the model has input parameters during the __init__ call.

  • train (bool) – Sets the model to training mode.

  • in_map (dict) – A mapping of the model’s input parameters and the training sample.

  • out_map (dict) – A mapping of the model’s output parameters to the truth training sample attributes.

  • loss_map (dict) – Returns the loss of each model output.

  • class_map (dict) – A map indicating which of the model’s output are used for classification modes.

  • code (Code) – Returns a Code object to rebuild the model object.

  • Epoch (int) – The epoch of the trained model.

  • KFold (int) – The kFold of the trained model.

  • Path (str) – The path of the model’s saved state.

  • RunName (str) – The run-name of the model.

  • model – Returns the original model type.

  • device (str) – The device that the model should be transferred to.

  • failure (bool) – Indicates whether the model’s code was not properly traced or initialized.

  • error_message (str) – Returns the error associated with the model.

  • KinematicMap (dict) – The kinematics to use for reconstructing the predicted mass from the nodes and topology.

Model Declarations (Example)

The Model wrapper class was introduced to do a preliminary scan of the model’s inputs and prevent the model from crashing due to missing features during training. To further improve performance during inference and training of a given model, the internal mapping only provides the model with the inputs it requires rather than injecting all features. To declare a specific input parameter, the pre-fixes G, N and E are used to indicate whether a Graph, Node or Edge feature is requested. A similar syntax is used to request truth features as inputs, by appending to the pre-fixes, _T.

class YourModel(...):

    def __init__(self):

        self.O_output = None
        self.L_output = "CEL"

        #... < some neural network initialization >...#

    def forward(self, edge_index, batch, G_<some graph feature>, N_<some node features>, E_<some edge feature>):

        # some logic...
        pass