This repository provides a library to compute and compare Visual Distribution of Neuron Activations (VDNAs). Visuals DNAs allow comparing datasets or images using distributions of neuron activations throughout layers of a feature extractor. They can be used as an alternative to FID and provide more holistic and granular comparisons.
VDNAs are generated by passing images to represent through a frozen pre-trained feature extractor network and monitoring activation values throughout the network. We can represent the images efficiently by fitting distributions such as histograms or Gaussians to the activations at each neuron. Because VDNAs keep neuron activations independent, we can compare VDNAs while focusing on neurons that are more relevant to the tasks at hand.
For example, we can compare VDNAs to find similar datasets based on attributes of interest selected through specific combinations of neurons.
We can also represent individual images and check their similarity to datasets or other images.
This work was introduced in our paper:
Visual DNA: Representing and Comparing Images using Distributions of Neuron Activations
Benjamin Ramtoula, Matthew Gadd, Paul Newman, Daniele de Martini
CVPR 2023
Please consider citing our paper if you make use of this project in your research.
@article{ramtoula2023vdna,
author = {Ramtoula, Benjamin and Gadd, Matthew and Newman, Paul and De Martini, Daniele},
title = {Visual DNA: Representing and Comparing Images using Distributions of Neuron Activations},
journal = {CVPR},
year = {2023},
}
You can also visit the project website.
Visual DNA can be installed using pip:
pip install vdna
Example of comparing two datasets using VDNAs using all default settings (histograms with 1000 bins, Mugs ViT-B/16 feature extractor, all neurons) from two dataset folders:
from vdna import VDNAProcessor, EMD
vdna_proc = VDNAProcessor()
vdna1 = vdna_proc.make_vdna(source="/path/to/dataset1")
vdna2 = vdna_proc.make_vdna(source="/path/to/dataset2")
emd = EMD(vdna1, vdna2)
This library provides tools to create VDNAs, save and load them, and compare them.
The general expected approach is to use the VDNAProcessor to create VDNA objects. These can be inspected and the distance between two VDNAs can be computed using provided functions.
See the documentation in the code for options such as the number of workers, batch size, or device.
Some of the settings allow selecting which feature extractor and distribution to use. See below for a list of supported options.
Example usage:
from vdna import VDNAProcessor
vdna_proc = VDNAProcessor()
vdna = vdna_proc.make_vdna(source="/path/to/dataset1", distribution_name="histogram-500", num_workers=16, batch_size=128)
We also support other input formats:
# From a list of NumPy arrays
vdna_from_numpy = vdna_proc.make_vdna(source=list_of_np_arrays)
# From a .txt file containing the paths to all images, one per line
vdna_from_txt = vdna_proc.make_vdna(source="path/to/index.txt")
# From a list of strings containing the paths to all images
vdna_from_txt = vdna_proc.make_vdna(source=["path/to/im1.png","path/to/im2.jpeg"])
You can save and load VDNAs using save and load_vdna_from_files. They both expect a path without any extension.
save will save two files in the given path and name, one .json with metadata and one .npz with distribution content.
load_vdna_from_files will load the two files and return a VDNA object.
Example:
# This will save files /path/to/save/vdna.json and /path/to/save/vdna.npz
vdna.save("/path/to/save/vdna")
# This will load the two files and return a VDNA object
from vdna import load_vdna_from_files
vdna = load_vdna_from_files("/path/to/save/vdna")
We also provide load_vdna_from_hub to load VDNAs directly from a HuggingFace Hub repository.
Once you have generated the VDNAs, you can access their distributions.
from vdna import load_vdna_from_files
# ----- Loading VDNAs -----
vdna1 = load_vdna_from_files("/path/to/save/vdna1")
vdna2 = load_vdna_from_files("/path/to/save/vdna2")
# ----- Checking feature extractor and distribution -----
print(f"vdna1 uses {vdna1.feature_extractor_name} as feature extractor and {vdna1.name} as distribution.")
# ----- Checking neurons used -----
print("List of layers and neurons in the VDNA:")
for layer_name in vdna1.neurons_list:
print(f"Layer {layer_name} has {vdna1.neurons_list[layer_name]} neurons")
# ----- Checking distribution values -----
all_neurons_hists = vdna1.get_all_neurons_dists()
print(f"We have {all_neurons_hists.shape[0]} neurons in the VDNA, with {all_neurons_hists.shape[1]} bins each.")
print(f"The highest value in a bin is {all_neurons_hists.max()}")
block_0_neurons_hists = vdna1.get_all_neurons_in_layer_dist(layer_name="block_0")
print(f"We have {block_0_neurons_hists.shape[0]} neurons in the VDNA using block_0 neurons, with {block_0_neurons_hists.shape[1]} bins each.")
print(f"The highest value in a bin is {block_0_neurons_hists.max()}")
specific_neuron_hist = vdna1.get_neuron_dist(layer_name="block_0", neuron_idx=42)
print(f"We have {specific_neuron_hist.shape[1]} bins in the histogram for neuron 42 in block_0.")
print(f"The highest value in a bin is {specific_neuron_hist.max()}")
We also provide distance functions that take in VDNAs to compare, for example with histogram-based VDNAs:
# ----- Comparing VDNAs -----
# Earth Mover's Distance used to compare histogram-based VDNAs
from vdna import load_vdna_from_files, EMD
vdna1 = load_vdna_from_files("/path/to/save/vdna1)
vdna2 = load_vdna_from_files("/path/to/save/vdna2)
print("EMD averaged over all neuron comparisons:")
print(EMD(vdna1, vdna2))
print("EMD averaged over all neurons of block_0:")
print(EMD(vdna1, vdna2, use_neurons_from_layer="block_0"))
print("EMD comparing neuron 42 of layer block_0")
print(EMD(vdna1, vdna2, use_neurons_from_layer="block_0", use_neuron_index=42))
print("Neuron-wise EMD comparisons as a dict:")
emd_neuron_wise = EMD(vdna1, vdna2, return_neuron_wise=True)
for layer in emd_neuron_wise:
print(f"EMD using neuron 42 of layer {layer} is {emd_neuron_wise[layer][42]}")
Visual DNAs can be constructed with different feature extractors and distributions. Here we detail supported options.
mugs_vit_base(default)mugs_vit_largevgg16inceptiondino_resnet50dino_vit_baserand_resnet50clip_im_vit_b16
histogram-{number of bins}(default ishistogram-1000):- This uses histograms for each neuron, as in the paper.
- In the paper, results are based on 1000 bins, although you might be fine using fewer.
- Can be compared with the Earth Mover's Distance:
from vdna import EMD.
gaussian:- This uses a Gaussian for each neuron, as in the paper.
- Can be compared with the neuron-wise Fréchet Distance:
from vdna import NFD.
layer-gaussian:- This fits a multivariate Gaussian for each selected layer of the feature extractor.
- Can be compared with the neuron-wise Fréchet Distance:
from vdna import NFD. - Can be compared with the layer-wise Fréchet Distance (to reproduce FID for example):
from vdna import FD.
activation-ranges:- This keeps track of minimum and maximum activation values for each neuron.
- Can be useful to get neuron activation ranges used for normalisation for histograms (see documentation about adding your feature extractor)
The commonly used Fréchet Inception Distance (FID) can be computed by creating VDNAs with the inception feature extractor and the layer-gaussian distribution.
It can then be computed using the FD function on layer block_3.
For example:
from vdna import VDNAProcessor, FD
vdna_proc = VDNAProcessor()
vdna1 = vdna_proc.make_vdna(source=dataset1_path, distribution_name="layer-gaussian", feat_extractor_name="inception")
vdna2 = vdna_proc.make_vdna(source=dataset2_path, distribution_name="layer-gaussian", feat_extractor_name="inception")
print("FID:")
print(FD(vdna1, vdna2, use_neurons_from_layer="block_3"))
Coming soon...
We are planning to pre-compute VDNAs for common datasets and store them online.
This repository was inspired and adapted using parts of the great clean-fid library.
Our model implementations and weights are adapted from the following repositories:
- https://github.com/sail-sg/mugs
- https://github.com/facebookresearch/swav
- https://github.com/facebookresearch/dino
- https://github.com/huggingface/pytorch-image-models
- https://github.com/isl-org/PhotorealismEnhancement
- https://github.com/bioinf-jku/TTUR
- https://github.com/openai/CLIP
- https://github.com/mseitzer/pytorch-fid
We are very thankful to the contributors of all these repositories for sharing their work.
The code in this repository is released under the MIT License.