GenLabel: Mixup Relabeling using Generative Models
Authors: Jy-Yong Sohn, Liang Shang, Hongxu Chen, Jaekyun Moon, Dimitris Papailiopoulos, Kangwook Lee
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Via theoretical and empirical analysis, we show that mixup, when used together with Gen Label, can effectively resolve the aforementioned phenomenon, improving the accuracy of mixup-trained model. ... We tested Gen Label on 133 low-dimensional real datasets in Open ML (Vanschoren et al., 2013). Our experimental results show that the suggested Gen Label helps mixup improve the accuracy in various low-dimensional datasets. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, University of Wisconsin, Madison, USA 2School of Electrical Engineering, Daejeon, KAIST. |
| Pseudocode | Yes | The pseudocode of Gen Label is provided in Algorithm 1. ... The pseudocode of this Gen Label variant (for the latent feature) is given in Algorithm 2. ... The pseudocode of this variant is given in Algorithm 3. |
| Open Source Code | Yes | The github repo for reproducible code is given in https://github.com/UW-Madison-Lee-Lab/Gen Label_official. |
| Open Datasets | Yes | We tested Gen Label on 133 low-dimensional real datasets in Open ML (Vanschoren et al., 2013). ... We also tested Gen Label on MNIST, CIFAR-10, CIFAR-100 and Tiny Image Net-200. ... The Circle and Moon datasets used in Table 1 are from scikit-learn (Pedregosa et al., 2011). |
| Dataset Splits | Yes | Here we choose the optimal mixing ratio γ using cross-validations. ... When we train mixup+Gen Label on Open ML datasets, we used a 6-fold cross-validation for choosing the best loss ratio γ {0.0, 0.2, 0.4, 0.6, 0.8, 1.0}. |
| Hardware Specification | Yes | We tested on NVIDIA Tesla V100 GPUs in Amazon Web Service (AWS) and local NVIDIA RTX2080 GPU machines. |
| Software Dependencies | No | All algorithms are implemented in Py Torch (Paszke et al., 2017) ... The code implemented in Tensor Flow (Abadi et al., 2015) ... scikit-learn (Pedregosa et al., 2011). The paper mentions the software tools used but does not provide specific version numbers for them. |
| Experiment Setup | Yes | For 2D and 3D cube datasets, we use a 3-layer fully connected network, which has 64 neurons in the first hidden layer and 128 neurons in the second hidden layer. ... For all the datasets, we use the SGD optimizer and the multi-step learning rate decay. ... The hyperparameters used in our experiments are summarized in Table 10, 11, 12, 13 and 14. |