OCDocker.OCScore.Dimensionality.future.losses module

Loss utilities for the future Autoencoder pipeline.

OCDocker.OCScore.Dimensionality.future.losses.contractive_penalty(embeddings, inputs)[source]

Compute contractive penalty (Jacobian norm).

Parameters:

  • embeddings (torch.Tensor) – Latent embeddings (N, D).

  • inputs (torch.Tensor) – Inputs with requires_grad=True (N, F).

Returns:

Scalar contractive penalty.

Return type:

torch.Tensor

OCDocker.OCScore.Dimensionality.future.losses.energy_loss(pred, target, loss_type='huber', huber_delta=1.0)[source]

Compute energy regression loss.

Parameters:

  • pred (torch.Tensor) – Predicted energies.

  • target (torch.Tensor) – Target energies.

  • loss_type (Literal["mse", "rmse", "mae", "huber"], optional) – Energy loss type, by default “huber”.

  • huber_delta (float, optional) – Delta parameter for Huber loss, by default 1.0.

Returns:

Scalar energy loss.

Return type:

torch.Tensor

OCDocker.OCScore.Dimensionality.future.losses.kl_divergence(mu, logvar)[source]

Compute KL divergence for VAE (mean over batch).

Parameters:

  • mu (torch.Tensor) – Latent mean tensor.

  • logvar (torch.Tensor) – Latent log-variance tensor.

Returns:

Scalar KL divergence loss.

Return type:

torch.Tensor

OCDocker.OCScore.Dimensionality.future.losses.reconstruction_loss(pred, target, loss_type='mse', huber_delta=1.0)[source]

Compute reconstruction loss between prediction and target.

Parameters:

  • pred (torch.Tensor) – Predicted reconstruction tensor.

  • target (torch.Tensor) – Target reconstruction tensor.

  • loss_type (Literal["mse", "rmse", "mae", "huber"], optional) – Reconstruction loss type, by default “mse”.

  • huber_delta (float, optional) – Delta parameter for Huber loss, by default 1.0.

Returns:

Scalar reconstruction loss.

Return type:

torch.Tensor