zea.models.lpips¶

LPIPS model for perceptual similarity.

To try this model, simply load one of the available presets:

>>> from zea.models.lpips import LPIPS

>>> model = LPIPS.from_preset("lpips")

Important

This is a zea implementation of the model. For the original paper and code, see here.

Zhang, Richard, et al. “The Unreasonable Effectiveness of Deep Features as a Perceptual Metric.” https://arxiv.org/abs/1801.03924

See also

Functions

`linear_model`()	Get the linear head model for LPIPS.
`perceptual_model`()	Get the VGG16 model for perceptual loss.

Classes

LPIPS(*args, **kwargs)

Learned Perceptual Image Patch Similarity (LPIPS) metric.

class zea.models.lpips.LPIPS(*args, **kwargs)[source]¶

Learned Perceptual Image Patch Similarity (LPIPS) metric.

Initialize the LPIPS model.

Exported weights using:: https://github.com/moono/lpips-tf2.x/blob/master/example_export_script/convert_to_tensorflow.py

Parameters:

net_type (str, optional) – Type of network to use. Defaults to “vgg”.
disable_checks (bool, optional) – Disable input checks. This is useful to allow tensorflow graph mode. Defaults to False.

call(inputs)[source]¶

Compute the LPIPS metric.

Parameters:

inputs (list) – List of two input images of shape [B, H, W, C] or [H, W, C]. Images should be in the range [-1, 1].

Returns:

LPIPS distance between the two images: of shape [B, ] or scalar if no batch dimension.

Return type:

Tensor

custom_load_weights(preset, **kwargs)[source]¶: Load the weights for the VGG and linear models.

static preprocess_input(image)[source]¶

Preprocess the input images

Parameters:

image (Tensor) – Input image tensor of shape [H, W, C] with optional batch dimension and values in the range [-1, 1].

Returns:

Preprocessed image tensor of shape [B, H, W, C]: and standardized values for VGG model.

Return type:

Tensor

zea.models.lpips.linear_model()[source]¶: Get the linear head model for LPIPS.

zea.models.lpips.perceptual_model()[source]¶: Get the VGG16 model for perceptual loss.