Can We Find Strong Lottery Tickets in Generative Models?
Authors: Sangyeop Yeo, Yoojin Jang, Jy-yong Sohn, Dongyoon Han, Jaejun Yoo
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results show that the discovered subnetwork can perform similarly or better than the trained dense model even when only 10% of the weights remain. |
| Researcher Affiliation | Collaboration | Sangyeop Yeo1, Yoojin Jang1, Jy-yong Sohn2, Dongyoon Han3, Jaejun Yoo1* 1LAIT, Ulsan National Institute of Science and Technology (UNIST) 2University of Wisconsin-Madison 3Naver AI Lab |
| Pseudocode | No | The paper provides mathematical equations (e.g., Equation 4) but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code and supplementary materials are publicly available at https://lait-cvlab.github.io/SLT-in-Generative-Models/. |
| Open Datasets | Yes | We use LSUN Bedroom (Yu et al. 2015), FFHQ (Karras, Laine, and Aila 2019), CIFAR10 (Krizhevsky, Hinton et al. 2009), Celeb A (Liu et al. 2015), and Baby Image Net (Kang, Shin, and Park 2022) datasets. |
| Dataset Splits | No | The paper mentions using '10,000 samples of real and generated images' for evaluation, but it does not specify explicit train/validation/test splits, percentages, or the methodology for partitioning the datasets for these specific purposes. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions using the 'Studio GAN' codebase and 'pytorch-studiogan' (https://github.com/postech-cvlab/pytorch-studiogan) but does not specify version numbers for PyTorch, Python, or other key software dependencies. |
| Experiment Setup | Yes | We test two pretrained models: (1) one using the default hyperparameter (solid red line) (Santos et al. 2019); (2) the other using a more optimized hyperparameter found from our experiments (solid blue line). Each dashed line shows the performance of the subnetwork obtained by our pruning method for each pretrained model with different target k. |