Skip to content

N2V Losses

Source

Noise2Void and related losses.

n2v_loss(manipulated_batch, original_batch, masks, *args)

N2V Loss function described in A Krull et al 2018.

Parameters:

  • manipulated_batch (Tensor) –

    Batch after manipulation function applied.

  • original_batch (Tensor) –

    Original images.

  • masks (Tensor) –

    Coordinates of changed pixels.

  • *args (Any, default: () ) –

    Additional arguments.

Returns:

  • Tensor –

    Loss value.

pn2v_loss(samples, labels, masks, noise_model)

Probabilistic N2V loss function described in A Krull et al., CVF (2019).

Parameters:

  • samples (Tensor) –

    Predicted pixel values from the network.

  • labels (Tensor) –

    Original pixel values.

  • masks (Tensor) –

    Coordinates of manipulated pixels.

  • noise_model (GaussianMixtureNoiseModel) –

    Noise model for computing likelihood.

Returns:

  • Tensor –

    Loss value.