Adversarial Schrödinger Bridge Matching

Authors: Nikita Gushchin, Daniil Selikhanovych, Sergei Kholkin, Evgeny Burnaev, Aleksandr Korotin

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our adversarial SB matching (ASBM) algorithm, which implements our D-IMF procedure on setups with Gaussian distributions (M4.1) for which we have closed form update formulas (M3.4) and real image data distributions (M4.2). We additionally provide results for an illustrative 2D example in Appendix C.1, results for the Colored MNIST dataset in Appendix C.3, and results on the standard SB benchmark in Appendix C.2.
Researcher Affiliation Academia Nikita Gushchin Skoltech Moscow, Russia n.gushchin@skoltech.ru Daniil Selikhanovych Skoltech Moscow, Russia selikhanovychdaniil@gmail.com Sergei Kholkin Skoltech Moscow, Russia s.kholkin@skoltech.ru Evgeny Burnaev Skoltech , AIRI Moscow, Russia e.burnaev@skoltech.ru Alexander Korotin Skoltech , AIRI Moscow, Russia a.korotin@skoltech.ru Equal contribution Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute
Pseudocode Yes Algorithm 1: Adversarial SB matching (ASBM). Input :number of intermediate steps N; initial process q0(x0, xt1, . . . , xt N , x1) accessible by samples; number of outer iteration K N; forward transitional density network {qθ(xtn|xtn 1)}N+1 n=1 ; backward transitional density network {qη(xtn 1|xtn)}N+1 n=1 ; Output :p0(x0) QN+1 n=1 qθ(xtn|xtn 1) p1(x1) QN+1 n=1 qη(xtn 1|xtn) p T (x0, xin, x1). for k = 0 to K 1 do Learn {qθ(xtn|xtn 1)}N+1 n=1 using 15 with q4k; Let q4k+1 be given by p0(x0) QN+1 n=1 qθ(xtn|xtn 1); Let q4k+2 be given by p W ϵ(xin|x0, x1)qθ(x0, x1); Learn {qη(xtn 1|xtn)}N+1 n=1 using 16 with q4k+2; Let q4k+3 be given by p1(x1) QN+1 n=1 qη(xtn 1|xtn); Let q4k+4 be given by p W ϵ(xin|x0, x1)qη(x0, x1);
Open Source Code Yes We provide the code at https://github.com/Daniil-Selikhanovych/ASBM.
Open Datasets Yes We evaluate our adversarial SB matching (ASBM) algorithm... on setups with Gaussian distributions (M4.1)... and real image data distributions (M4.2). We additionally provide results for an illustrative 2D example in Appendix C.1, results for the Colored MNIST dataset in Appendix C.3, and results on the standard SB benchmark in Appendix C.2. ... Celeba dataset [33] ... The SB mixtures benchmark proposed by [13, M4]...
Dataset Splits No For Celeb A: 'We use 10% of male and female images as the test set for evaluation.' For Colored MNIST: 'we use the default MNIST train/test split'. No explicit validation split information is provided for any dataset.
Hardware Specification Yes The most time challenging experiment on Celeb A runs for approximately 7 days on 1 GPUs A100. Experiment with Colored MNIST takes less then 2 days of training on GPU A100.
Software Dependencies No The paper mentions 'The code for our algorithm and all experiments with it is written in Pytorch' and 'Discrete Markovian Projection is conducted using the DD-GAN code [53]', 'Adam optimizer [22]', but does not specify version numbers for PyTorch, DD-GAN, or other libraries.
Experiment Setup Yes The hyperparameters which we use in the experiments are summarized in Table 4. Experiment Start couping q0(x0, x1) D-IMF outer iters D-IMF=0 grad updates D-IMF grad updates N Batch Size D/G opt ratio EMA decay Lr G Lr D 2D Toy Ind 20 400000 40000 3 512 1:1 0.999 1e-4 1e-4 SB Bench Ind 2 133000 67000 31 128 3:1 0.999 1e-4 1e-4 C-MNIST MB 3 100000 50000 3 64 1:1 0.999 1.25e-4 1.6e-4 Celeb A MB 5 1000000 40000 3 32 1:1 0.9999 1.25e-4 1.6e-4