metamod.ANN_pt module

Custom ANN definition using TensorFlow.

This module contains the definition of an ANN comptable with the SMT Toolbox.

metamod.ANN_pt.activations

Available activation functions.

Type:dict
metamod.ANN_pt.initializers

Available kernel initializers.

Type:dict

Notes

Keras tuner logs are not stored in data due to max path length issues in kerastuner.

See also

PyTorch documentation - https://pytorch.org/docs/stable/index.html

class metamod.ANN_pt.ANN_pt(**kwargs)

Bases: metamod.ANN.ANN_base

ANN class.

name

Name of the surrogate model.

Type:str
model

The model of the ANN.

early_stop

Number of training epchs before early stopping.

Type:int
build_hypermodel(hp)

Build the hypermodel of the ANN with Keras Tuner hyperparameters.

Parameters:hp (kerastuner.engine.hyperparameters.HyperParameters) – Hyperparameters.
Returns:The model of the ANN.
Return type:model ()

Notes

No kernel regularizers implemented so far.

See also

Keras Tuner documentation: https://keras-team.github.io/keras-tuner/

pretrain(inputs, outputs, iteration)

Optimize the hyperparameters of the ANN.

Parameters:
  • inputs (np.array) – All input data.
  • outputs (np.array) – All output data.
  • iteration (int) – Iteration number.
Returns:

Optimal hyperparameters.
Return type:

best_hp (kerastuner.engine.hyperparameters.HyperParameters)

Notes

Optimization objective fixed on val_loss.

save()

Save the ANN into an external file.

class metamod.ANN_pt.SparseModel(neurons_hyp, activation_hyp, kernel_regularizer, in_dim, layers_hyp, out_dim, init, bias_init)

Bases: torch.nn.modules.module.Module

Defines an ANN with a subnetwork for each output.

activation

Activation function.

Type:function
subnetworks

Neural network.

early_stopping(history, metric, tolerance, patience)

Check whether the early stopping condition has been met.

Parameters:
  • history (metamod.ANN_pt.TrainHistory) – Metrics history during the training.
  • metric (str) – Early stopping decision metric.
  • tolerance (float) – Early stopping tolerance.
  • patience (int) – Early stopping patience.
Returns:

Whether to early stop the training.
Return type:

stop (bool)
fit(epochs, train_in, train_out, test_in=None, test_out=None, optimizing=False)

Train the ANN.

Parameters:
  • epochs (int) –
  • train_in (torch.Tensor) – Train input data.
  • train_out (torch.Tensor) – Train output data.
  • test_in (torch.Tensor) – Test input data.
  • test_out (torch.Tensor) – Test output data.
  • optimizing (bool) – Whether this is an optimization run.
Returns:

Metrics history during the training.
Return type:

history (metamod.ANN_pt.TrainHistory)
forward(x)

Forward-pass throught the network.

Parameters:x (torch.Tensor) – Input tensor.
Returns:Output tensor.
Return type:out (torch.Tensor)
class metamod.ANN_pt.TrainHistory

Bases: object

This class stored the metrics history.

history

Stores training and validation losses.

Type:dict
store(loss, loss_eval=None)
metamod.ANN_pt.swish(x)

Swish activation function.

Parameters:x (torch.Tensor) – Weighted inputs.
Returns:Neuron’s activations.
Return type:activation (torch.Tensor)