Generative Modeling with Optimal Transport Maps
Authors: Litu Rout, Alexander Korotin, Evgeny Burnaev
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the algorithm on image generation and unpaired image restoration tasks. In particular, we consider denoising, colorization, and inpainting, where the optimality of the restoration map is a desired attribute, since the output (restored) image is expected to be close to the input (degraded) one. 5 EXPERIMENTS We evaluate our algorithm in generative modeling of the data distribution from a noise (M5.1) and unpaired image restoration task (M5.2). Technical details are given in Appendix B. Additionally, in Appendix B.4 we test our method on toy 2D datasets and evaluate it on the Wasserstein-2 benchmark (Korotin et al., 2021b) in Appendix B.2. |
| Researcher Affiliation | Collaboration | Litu Rout Space Applications Centre Indian Space Research Organisation lr@sac.isro.gov.in Alexander Korotin Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute (AIRI) a.korotin@skoltech.ru Evgeny Burnaev Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute (AIRI) e.burnaev@skoltech.ru |
| Pseudocode | Yes | Algorithm 1: Learning the optimal transport map between unequal dimensions. |
| Open Source Code | Yes | The Py Torch source code is provided at https://github.com/Litu Rout/Optimal Transport Modeling |
| Open Datasets | Yes | We test our method on MNIST 32x32 (Le Cun et al., 1998), CIFAR10 32x32 (Krizhevsky et al., 2009), and Celeb A 64x64 (Liu et al., 2015) image datasets. Anime128x1283. This dataset consists of 500000 high resolution images. We resize the cropped images as in Celeb A to 128x128, i.e. y R128 128 3. Here, KG = 5, Kψ = 1, λ = 0.01, learning rate 2 10 4, batch size 16, and betas=(0, 0.9). Anime: https://www.kaggle.com/reitanaka/alignedanimefaces |
| Dataset Splits | No | The paper describes a specific partitioning for unpaired image restoration: 'We split the dataset in 3 parts A, B, C containing 90K, 90K, 22K samples respectively. To each image we apply the degradation transform (decolorization, noising or occlusion) and obtain the degraded dataset containing of 3 respective parts A, B, C. For unpaired training we use part A of degraded and part B of clean images. For testing, we use parts C.' However, it does not provide explicit train/validation/test splits with percentages or counts for general reproducibility beyond this specific setup, nor does it explicitly mention a dedicated validation split for hyperparameter tuning. |
| Hardware Specification | Yes | All the experiments are conducted on 2 V100 GPUs. |
| Software Dependencies | No | The paper states 'We use the Py Torch framework' and mentions 'Adam (Kingma & Ba, 2014)' as the optimizer, but does not specify exact version numbers for PyTorch or any other software libraries used. |
| Experiment Setup | Yes | MNIST (Le Cun et al., 1998). On MNIST, we use x R192 and y R32 32. The batch size is 64, learning rate 2 10 4, optimizer Adam (Kingma & Ba, 2014) with betas (0, 0.9), gradient optimality coefficient λ = 10, and the number of training epochs T = 30. CIFAR10 (Krizhevsky et al., 2009). We use all 50000 samples while training. The latent vector x R192 and y R32 32 3, batch size 64, λ = 10, k G = 1, kψ = 1, T = 1000, Adam optimizer with betas (0, 0.9), and learning rate 2 10 4 for G and 1 10 3 for ψ. |