Wasserstein Iterative Networks for Barycenter Estimation

Authors: Alexander Korotin, Vage Egiazarian, Lingxiao Li, Evgeny Burnaev

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

Reproducibility Variable Result LLM Response
Research Type Experimental We develop a novel iterative algorithm (M4) for estimating Wasserstein-2 barycenters... We construct the Ave, celeba! (averaging celebrity faces, M5) dataset... for large-scale quantitative evaluation... We evaluate our iterative algorithm 1 and a recent state-of-the-art variational [SCW2B] by [19] on Ave, celeba! dataset... for quantitative evaluation we use FID score [23] computed on 200K generated samples w.r.t. the original Celeb A dataset, see Table 1. Our method drastically outperforms [SCW2B].
Researcher Affiliation Academia Alexander Korotin Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute Moscow, Russia a.korotin@skoltech.ru Vage Egiazarian Skolkovo Institute of Science and Technology Moscow, Russia vage.egiazarian@skoltech.ru Lingxiao Li Massachusetts Institute of Technology Cambridge, Massachusetts, USA lingxiao@mit.edu Evgeny Burnaev Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute Moscow, Russia e.burnaev@skoltech.ru
Pseudocode Yes Algorithm 1: Wasserstein Iterative Networks (WIN) for Barycenter Estimation
Open Source Code Yes The code1 is written on the Py Torch and includes the script for producing Ave, celeba! dataset. 1https://github.com/iamalexkorotin/Wasserstein Iterative Networks
Open Datasets Yes We use Celeb A 64 64 faces dataset [36] as the basis for our Ave, celeba! dataset... 50K Shoes [68], 138K Amazon Handbags and 90K Fruits [45].
Dataset Splits No The paper describes creating the Ave, celeba! dataset by splitting images into 3 parts for barycenter calculation, and mentions using mini-batches for training, but does not provide explicit train/validation/test splits for the generative model (Gξ) itself.
Hardware Specification Yes The experiments are conducted on 4 GPU GTX 1080ti.
Software Dependencies No The paper mentions 'Py Torch' but does not specify a version number or other software dependencies with versions.
Experiment Setup Yes In practice, we choose KG = 50 as it empirically works better... Input :latent S and input P1, . . . , PN measures; weights α1, . . . , αN > 0 (PN n=1 αn = 1); number of iters per network: KG, KT , Kv; generator Gξ : RH RD; mapping networks Tθ1, . . . , TθN : RD RD; potentials vω1, . . . , vωN : RD R; regression loss ℓ: RD RD R+;... In Lemma 3, we set N = 3, M = 2, β1 = β2 = 1/2, w1 = w2 = 1/2, (γl) = 1 0 0 0 1 0 , (γr) = 0 1 0 0 0 1 which yields weights (α1, α2, α3) = (1/4, 1/2, 1/4).