evox.problems.neuroevolution.supervised_learning

evox.problems.neuroevolution.supervised_learning#

Module Contents#

Classes#

SupervisedLearningProblem

The supervised learning problem to test a model’s parameters or a batch of parameters with given data and criterion.

Data#

__all__

API#

evox.problems.neuroevolution.supervised_learning.__all__#

[‘SupervisedLearningProblem’]

class evox.problems.neuroevolution.supervised_learning.SupervisedLearningProblem(model: torch.nn.Module, data_loader: torch.utils.data.DataLoader, criterion: torch.nn.Module, n_batch_per_eval: int = 1, pop_size: int | None = None, device: torch.device | None = None, reduction: str = 'mean')[source]#

Bases: evox.core.Problem

The supervised learning problem to test a model’s parameters or a batch of parameters with given data and criterion.

Initialization

Initialize the SupervisedLearningProblem.

Parameters:

  • model – The neural network model whose parameters need to be evaluated.

  • data_loader – The data loader providing the dataset for evaluation.

  • criterion – The loss function used to evaluate the parameters’ performance.

  • n_batch_per_eval – The number of batches to be evaluated in a single evaluation. When set to -1, will go through the whole dataset. Defaults to 1.

  • pop_size – The size of the population (batch size of the parameters) to be evaluated. Defaults to None for single-run mode.

  • device – The device to run the computations on. Defaults to the current default device.

  • reduction – The reduction method for the criterion. ‘mean’ | ‘sum’. Defaults to “mean”.

Raises:

RuntimeError – If the data loader contains no items.

Warning

This problem does NOT support HPO wrapper (problems.hpo_wrapper.HPOProblemWrapper), i.e., the workflow containing this problem CANNOT be vmapped.

_vmap_forward_pass(model_state: Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]], data: Tuple[torch.Tensor, torch.Tensor])[source]#
_forward_pass(model_state: Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]], data: Tuple[torch.Tensor, torch.Tensor])[source]#
_vmap_evaluate(pop_params: Dict[str, torch.nn.Parameter])[source]#
_single_evaluate(params: Dict[str, torch.nn.Parameter])[source]#
evaluate(pop_params: Dict[str, torch.nn.Parameter]) → torch.Tensor[source]#

Evaluate the fitness of a population (batch) of model parameters.

Parameters:

pop_params – A dictionary of parameters where each key is a parameter name and each value is a tensor of shape (batch_size, *param_shape) representing the batched parameters of batched models.

Returns:

A tensor of shape (batch_size,) containing the fitness of each sample in the population.

Contents