Neural Monge Map estimation and its applications
Authors: Jiaojiao Fan, Shu Liu, Shaojun Ma, Hao-Min Zhou, Yongxin Chen
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The performance of our algorithms is demonstrated through a series of experiments with both synthetic and realistic data, including text-to-image generation, class-preserving map, and image inpainting tasks. |
| Researcher Affiliation | Academia | Jiaojiao Fan EMAIL Georgia Institute of Technology Shu Liu EMAIL University of California, Los Angeles Shaojun Ma EMAIL Georgia Institute of Technology Hao-Min Zhou EMAIL Georgia Institute of Technology Yongxin Chen EMAIL Georgia Institute of Technology |
| Pseudocode | Yes | Algorithm 1 Computing optimal Monge map from ρa to ρb 1: Input: Marginal distributions ρa and ρb, Batch size B, Cost function c(x, y). 2: Initialize Tθ, fη. 3: for K steps do 4: Sample {Xk}B k=1 ρa. Sample {Yk}B k=1 ρb. 5: Update θ to decrease (6) for K1 steps. 6: Update η to increase (6) for K2 steps. 7: end for 8: Output: The transport map Tθ. |
| Open Source Code | No | The paper mentions using "POT implementation (Flamary et al., 2021) of Perrot et al. (2016)" but does not explicitly state that the authors are releasing their own code for the methodology described in this paper, nor does it provide a direct link to a repository. |
| Open Datasets | Yes | We evaluate our algorithm on two datasets: Laion art and Conceptual Captions 3M (CC-3M). ... We compare our algorithm with Asadulaev et al. (2022) on NIST (Le Cun & Cortes, 2005; Xiao et al., 2017; Clanuwat et al., 2018) datasets, ... We conduct the experiments on Celeb A 64 64 and 128 128 datasets (Liu et al., 2015). ... The samples used in our experiment are generated from the licensed dataset from Doxsey-Whitfield et al. (2015). |
| Dataset Splits | Yes | For each dataset, we let the source and the target distribution contain 0.3M data respectively, and take the rest of dataset as the test data. |
| Hardware Specification | Yes | On NVIDIA RTX A6000 (48GB), the training time of each experiment is 21 hours. ... On NVIDIA RTX A6000 (48GB), the training time of Celeb A64 experiment is 10 hours and the time of Celeb A128 is 45 hours. |
| Software Dependencies | No | The paper mentions software like "Adam method Kingma & Ba (2014)", "Py Torch", "CLIP model", "DALL E2 diffusion decoder", and "POT library Flamary et al. (2021)" but does not provide specific version numbers for these components. |
| Experiment Setup | Yes | The learning rate is 10 3. The number of iterations K = 12000. ... The training batch size B = 2000. We set K = 8000, K1 = 8, K2 = 6. The learning rate for both θ and η equals 10 4. ... The batch size is 225. The numbers of loop iterations are K1 = 10, K2 = 1. We use the learning rates 10 4, Adam (Kingma & Ba, 2014) optimizer with weight decay coefficient 0.0602. We train the networks for 110 epochs. |