E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation

Authors: Yifan Gong, Zheng Zhan, Qing Jin, Yanyu Li, Yerlan Idelbayev, Xian Liu, Andrey Zharkov, Kfir Aberman, Sergey Tulyakov, Yanzhi Wang, Jian Ren

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

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments In this section, we provide the detailed experimental settings and results of our proposed method.
Researcher Affiliation Collaboration Yifan Gong * 1 2 Zheng Zhan * 2 Qing Jin 1 Yanyu Li 1 2 Yerlan Idelbayev 2 Xian Liu 1 Andrey Zharkov 1 Kfir Aberman 1 Sergey Tulyakov 1 Yanzhi Wang 2 Jian Ren 1 *Equal contribution, work is done during Yifan s internship at Snap Inc. 1Snap Inc. 2Northeastern University. Correspondence to: Jian Ren <jren@snapchat.com>.
Pseudocode Yes The overall algorithm is described in Algorithm 1 in Sec. A in the Appendix.
Open Source Code No Project Page: https://yifanfanfanfan.github.io/e2gan/. The project page is mentioned, but there is no explicit statement that the source code for the methodology described in the paper is available there, nor is it a direct link to a code repository.
Open Datasets Yes We verify our method on 1,000 images from FFHQ dataset (Karras et al., 2019) and Flickr Scenery dataset (Cheng et al., 2022) with image resolution as 256 256.
Dataset Splits Yes To perform training and evaluation of GAN models, we divide the image pairs from each target concept into training/validation/test subsets with the ratio as 80%/10%/10%.
Hardware Specification Yes The training and training time measurements are conducted on one NVIDIA H100 GPU with 80 GB memory.
Software Dependencies No The paper mentions using the Adam solver, but does not provide specific version numbers for software libraries or environments like Python, PyTorch, or CUDA.
Experiment Setup Yes The training is conducted from an initial learning rate of 2e-4 with mini-batch SGD using Adam solver (Kingma & Ba, 2014). The total training epochs is set to 100 for E2GAN, and 200 for pix2pix (Isola et al., 2017) and pix2pix-zero-distilled (Parmar et al., 2023) for them to converge well.