normalizing_flow

Standard model class(es).

class graphnet.models.normalizing_flow.NormalizingFlow(*args, **kwargs)[source]

Bases: EasySyntax

A model for building (conditional) normalizing flows in GraphNeT.

This model relies on jammy_flows for building and evaluating normalizing flows. https://thoglu.github.io/jammy_flows/usage/introduction.html for details.

Build NormalizingFlow to learn (conditional) normalizing flows.

NormalizingFlow is able to build, train and evaluate a wide suite of normalizing flows. Instead of optimizing a loss function, flows minimize a learned pdf of your data, providing you with a posterior distribution for every example instead of point-like predictions.

NormalizingFlow can be conditioned on existing fields in the DataRepresentation or latent representations from Models.

NormalizingFlow is built upon https://github.com/thoglu/jammy_flows, and we refer to their documentation for details on the flows.

Parameters:

  • graph_definition (GraphDefinition) – The GraphDefinition to train the model on.

  • target_labels (str) – Name of target(s) to learn the pdf of.

  • backbone (Optional[GNN], default: None) – Architecture used to produce latent representations of

  • conditioned. (the input data on which the pdf will be)

  • None. (Defaults to)

  • condition_on (Union[List[str], str, None], default: None) – List of fields in Data objects to condition the

  • None.

  • flow_layers (str, default: 'gggt') – A string defining the flow layers.

  • https (See) – //thoglu.github.io/jammy_flows/usage/introduction.html

  • "gggt". (for details. Defaults to)

  • optimizer_class (Type[Optimizer], default: <class 'torch.optim.adam.Adam'>) – Optimizer to use. Defaults to Adam.

  • optimizer_kwargs (Optional[Dict], default: None) – Optimzier arguments. Defaults to None.

  • scheduler_class (Optional[type], default: None) – Learning rate scheduler to use. Defaults to None.

  • scheduler_kwargs (Optional[Dict], default: None) – Arguments to learning rate scheduler.

  • None.

  • scheduler_config (Optional[Dict], default: None) – Defaults to None.

  • args (Any)

  • kwargs (Any)

Raises:

ValueError – if both backbone and condition_on is specified.

Return type:

object

forward(data)[source]

Forward pass, chaining model components.

Return type:

Tensor

Parameters:

data (Data | List[Data])

shared_step(batch, batch_idx)[source]

Perform shared step.

Applies the forward pass and the following loss calculation, shared between the training and validation step.

Return type:

Tensor

Parameters:

  • batch (List[Data])

  • batch_idx (int)

validate_tasks()[source]

Verify that self._tasks contain compatible elements.

Return type:

None