StrWAEs to Invariant Representations

Authors: Hyunjong Lee, Yedarm Seong, Sungdong Lee, Joong-Ho Won

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimented Str WAEs with various real-world datasets. The generative model in Section 3.2.1 is applied for semi-supervised learning and conditional generation on the MNIST and SVHN (Netzer et al., 2011) datasets. The models in Sections 3.2.2 and 3.2.3 are used for learning conditional generation on the VGGFace2 datasets (Cao et al., 2018) and invariant representation on the Extended Yale B dataset (Georghiades et al., 2001; Lee et al., 2005).
Researcher Affiliation Academia 1Department of Statistics, Seoul National University, Seoul, Korea 2Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul, Korea 3Department of Medicine, Yong Loo Lin School of Medicine, National University of Singapore, Singapore.
Pseudocode Yes Algorithm 1 Traning Str WAEs in Sections 3.2.1 to 3.2.3
Open Source Code Yes Code available at https://github.com/comp-stat/StrWAE
Open Datasets Yes We experimented Str WAEs with various real-world datasets. The generative model in Section 3.2.1 is applied for semi-supervised learning and conditional generation on the MNIST and SVHN (Netzer et al., 2011) datasets. The models in Sections 3.2.2 and 3.2.3 are used for learning conditional generation on the VGGFace2 datasets (Cao et al., 2018) and invariant representation on the Extended Yale B dataset (Georghiades et al., 2001; Lee et al., 2005).
Dataset Splits No The paper provides details for training and test splits for several datasets (e.g., 'For each subject, the pictures of the person are split into training and test data with a fixed ratio, resulting in 1,664 and 750 images for the training and test respectively.' for Extended Yale B, and '9:1 train-test split' for Mini Speech Recognition), but does not explicitly specify a validation dataset split.
Hardware Specification Yes We trained the networks with Intel Xeon CPU Silver 4114 @ 2.20GHz processors and Nvidia Titan V GPUs with 12GB memory.
Software Dependencies Yes All the implementations were based on Python 3.11, Py Torch 2.1.1, and CUDA 12.1.
Experiment Setup Yes We set the hyper-parameters as follows: for the MNIST, λ1 = 100, λ2 = 100, µ1 = 500, and µ2 = 0 for the SVHN, λ1 = 10, λ2 = 10, µ1 = 1000, and µ2 = 0. We trained the model end-to-end with 500 and 200 epochs, respectively.