image-diversity 0.1.6

A package for computing entropy-based diversity measures on image sets

MEASURING DIVERSITY OF IMAGE SETS

Assess a set of image's diversity by computing an approximation of the Entropy of the images' mappings into a pre-trained network latents.

Installation

Install from pypi:

pip install image-diversity

Usage (console)

To compute the Truncated CLIP Entropy (TCE) of a set of images in a directory, run

python3 -m image_diversity <path/to/dir>

To compute the Truncated Inception Entropy (TIE), change the div_type option to TIE

python3 -m image_diversity <path/to/dir> --div_type TIE

Arguments

python3 -m image_diversity <path/to/dir>
--div_type <str>        Encoding network: TCE (default) or TIE
--n_eigs <int>          Number of eigenvalues for truncation (default: 20)
--device <str>          Device to use. Automatic by default
--batch_size <int>      Batch size for data loading (default: 16)

Usage (script)

Computing Truncated CLIP Entropy (TCE)

from image_diversity import ClipMetrics

clip_metrics = ClipMetrics()
tce = clip_metrics.tce("path/to/img/dir")

Computing Truncated Inception Entropy (TIE)

from image_diversity import InceptionMetrics

inception_metrics = InceptionMetrics()
tie = inception_metrics.tie("path/to/img/dir")

Inception implementation based on pytorch-fid

CLIP implementation based on CLIP