Measuring dissimilarity with diffeomorphism invariance

Authors: Théophile Cantelobre, Carlo Ciliberto, Benjamin Guedj, Alessandro Rudi

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical experiments support the merits of DID. and We support our theoretical claims with numerical experiments on images. and This section investigates the empirical performance of DID.
Researcher Affiliation Academia DI ENS, Ecole normale sup erieure, Universit e PSL, CNRS, Inria, Paris, France, Inria London, UK, Centre for AI, Department of Computer Science, University College London, UK, Inria, Lille Nord Europe Research Centre, Lille, France.
Pseudocode No The paper describes algorithmic steps in paragraph form, 'An efficient algorithm to compute it consists in (1) first computing the matrices b A and b B, then the inverse b C = ( b B b B + λ) 1 and finally compute the largest eigenvalue of b A b C b A via a power iteration method', but does not include a structured pseudocode or algorithm block.
Open Source Code Yes The Python source code used for the experiments presented here is freely available at https://github.com/theophilec/diffy, depends on Numpy and Pytorch and supports using GPUs.
Open Datasets Yes We rely on images from Imagenet (more precisely from the Imagenette subset), example images from the Matlab software (peppers), and finally images taken with our personal devices for illustrations (raccon, flowers, objects). and Imagenette We use images from the Imagenet dataset for our experiments. We use the subset called Imagenette, available at: https://github.com/fastai/imagenette.
Dataset Splits No The paper mentions using Imagenette and other images but does not provide specific details on training, validation, or test splits, such as percentages, counts, or a cross-validation setup.
Hardware Specification No The paper vaguely mentions 'supports using GPUs' but does not provide specific hardware details such as exact GPU models, CPU models, or detailed computer specifications used for experiments.
Software Dependencies Yes We conduct our experiments using Pytorch version 1.9.0 (torchvision 0.10.0) and Numpy 1.20.1 (but our implementation does not require any special functions so should generalize to any recent version). We also make use of Kornia version 0.5.7 and PIL version 8.2.0 for IO.
Experiment Setup Yes DID has parameters: MX = 100, MY = 163, σ = 1/6 and a = 5. and In practice we normalize b A and b B by their operator norms b A op and b B op. and We choose µ to be a Blackman window