Learned Perceptual Image Patch Similarity (LPIPS)

Module Interface

class torchmetrics.image.lpip.LearnedPerceptualImagePatchSimilarity(net_type='alex', reduction='mean', normalize=False, **kwargs)

The Learned Perceptual Image Patch Similarity (LPIPS_) calculates perceptual similarity between two images.

LPIPS essentially computes the similarity between the activations of two image patches for some pre-defined network. This measure has been shown to match human perception well. A low LPIPS score means that image patches are perceptual similar.

Both input image patches are expected to have shape (N, 3, H, W). The minimum size of H, W depends on the chosen backbone (see net_type arg).

Hint

Using this metrics requires you to have torchvision package installed. Either install as pip install torchmetrics[image] or pip install torchvision.

As input to forward and update the metric accepts the following input

• img1 (Tensor): tensor with images of shape (N, 3, H, W)

• img2 (Tensor): tensor with images of shape (N, 3, H, W)

As output of forward and compute the metric returns the following output

• lpips (Tensor): returns float scalar tensor with average LPIPS value over samples

Parameters:

• net_type (Literal['vgg', 'alex', 'squeeze']) – str indicating backbone network type to use. Choose between ‘alex’, ‘vgg’ or ‘squeeze’

• reduction (Literal['sum', 'mean']) – str indicating how to reduce over the batch dimension. Choose between ‘sum’ or ‘mean’.

• normalize (bool) – by default this is False meaning that the input is expected to be in the [-1,1] range. If set to True will instead expect input to be in the [0,1] range.

• kwargs (Any) – Additional keyword arguments, see Advanced metric settings for more info.

Raises:

• ModuleNotFoundError – If torchvision package is not installed

• ValueError – If net_type is not one of "vgg", "alex" or "squeeze"

• ValueError – If reduction is not one of "mean" or "sum"

Example

>>> from torch import rand
>>> from torchmetrics.image.lpip import LearnedPerceptualImagePatchSimilarity
>>> lpips = LearnedPerceptualImagePatchSimilarity(net_type='squeeze')
>>> # LPIPS needs the images to be in the [-1, 1] range.
>>> img1 = (rand(10, 3, 100, 100) * 2) - 1
>>> img2 = (rand(10, 3, 100, 100) * 2) - 1
>>> lpips(img1, img2)
tensor(0.1024)

plot(val=None, ax=None)

Plot a single or multiple values from the metric.

Parameters:

• val (Union[Tensor, Sequence[Tensor], None]) – Either a single result from calling metric.forward or metric.compute or a list of these results. If no value is provided, will automatically call metric.compute and plot that result.

• ax (Optional[Axes]) – An matplotlib axis object. If provided will add plot to that axis

Return type:

Tuple[Figure, Union[Axes, ndarray]]

Returns:

Figure and Axes object

Raises:

ModuleNotFoundError – If matplotlib is not installed

>>> # Example plotting a single value
>>> import torch
>>> from torchmetrics.image.lpip import LearnedPerceptualImagePatchSimilarity
>>> metric = LearnedPerceptualImagePatchSimilarity(net_type='squeeze')
>>> metric.update(torch.rand(10, 3, 100, 100), torch.rand(10, 3, 100, 100))
>>> fig_, ax_ = metric.plot()

>>> # Example plotting multiple values
>>> import torch
>>> from torchmetrics.image.lpip import LearnedPerceptualImagePatchSimilarity
>>> metric = LearnedPerceptualImagePatchSimilarity(net_type='squeeze')
>>> values = [ ]
>>> for _ in range(3):
...     values.append(metric(torch.rand(10, 3, 100, 100), torch.rand(10, 3, 100, 100)))
>>> fig_, ax_ = metric.plot(values)

Functional Interface

torchmetrics.functional.image.learned_perceptual_image_patch_similarity(img1, img2, net_type='alex', reduction='mean', normalize=False)

The Learned Perceptual Image Patch Similarity (LPIPS_) calculates perceptual similarity between two images.

LPIPS essentially computes the similarity between the activations of two image patches for some pre-defined network. This measure has been shown to match human perception well. A low LPIPS score means that image patches are perceptual similar.

Both input image patches are expected to have shape (N, 3, H, W). The minimum size of H, W depends on the chosen backbone (see net_type arg).

Parameters:

• img1 (Tensor) – first set of images

• img2 (Tensor) – second set of images

• net_type (Literal['alex', 'vgg', 'squeeze']) – str indicating backbone network type to use. Choose between ‘alex’, ‘vgg’ or ‘squeeze’

• reduction (Literal['sum', 'mean']) – str indicating how to reduce over the batch dimension. Choose between ‘sum’ or ‘mean’.

• normalize (bool) – by default this is False meaning that the input is expected to be in the [-1,1] range. If set to True will instead expect input to be in the [0,1] range.

Return type:

Tensor

Example

>>> from torch import rand
>>> from torchmetrics.functional.image.lpips import learned_perceptual_image_patch_similarity
>>> img1 = (rand(10, 3, 100, 100) * 2) - 1
>>> img2 = (rand(10, 3, 100, 100) * 2) - 1
>>> learned_perceptual_image_patch_similarity(img1, img2, net_type='squeeze')
tensor(0.1005)