LPIPS

class deepinv.loss.metric.LPIPS(device='cpu', check_input_range=False, **kwargs)[source]

Bases: Metric

Learned Perceptual Image Patch Similarity (LPIPS) metric.

Calculates the LPIPS \(\text{LPIPS}(\hat{x},x)\) where \(\hat{x}=\inverse{y}\).

Computes the perceptual similarity between two images, based on a pre-trained deep neural network. Uses implementation from pyiqa.

Note

By default, no reduction is performed in the batch dimension.

Example:

>>> from deepinv.utils.demo import get_image_url, load_url_image
>>> from deepinv.loss.metric import LPIPS
>>> ();m = LPIPS();() 
(...)
>>> x = load_url_image(get_image_url("celeba_example.jpg"), img_size=128)
>>> x_net = x - 0.01
>>> m(x_net, x) 
tensor([...])
Parameters:

  • device (str) – device to use for the metric computation. Default: ‘cpu’.

  • complex_abs (bool) – perform complex magnitude before passing data to metric function. If True, the data must either be of complex dtype or have size 2 in the channel dimension (usually the second dimension after batch).

  • reduction (str) – a method to reduce metric score over individual batch scores. mean: takes the mean, sum takes the sum, none or None no reduction will be applied (default).

  • norm_inputs (str) – normalize images before passing to metric. l2``normalizes by L2 spatial norm, ``min_max normalizes by min and max of each input.

  • check_input_range (bool) – if True, pyiqa will raise error if inputs aren’t in the appropriate range [0, 1].

metric(x_net, x, *args, **kwargs)[source]

Calculate metric on data.

Override this function to implement your own metric. Always include args and kwargs arguments.

Parameters:

  • x_net (torch.Tensor) – Reconstructed image \(\hat{x}=\inverse{y}\) of shape (B, ...) or (B, C, ...).

  • x (torch.Tensor) – Reference image \(x\) (optional) of shape (B, ...) or (B, C, ...).

Return torch.Tensor:

calculated metric, the tensor size might be (1,) or (B,).