Semi-Supervised StyleGAN for Disentanglement Learning

Authors: Weili Nie, Tero Karras, Animesh Garg, Shoubhik Debnath, Anjul Patney, Ankit Patel, Animashree Anandkumar

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We investigate the impact of limited supervision and find that using only 0.25% 2.5% of labeled data is sufficient for good disentanglement on both synthetic and real datasets. We propose new metrics to quantify generator controllability, and observe there may exist a crucial trade-off between disentangled representation learning and controllable generation.3.1. Experimental Setup3.2. Key Results
Researcher Affiliation Collaboration 1Rice University ( Work done as a part of internship at Nvidia) 2Nvidia 3University of Toronto 4California Institute of Technology.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described in this paper.
Open Datasets Yes We consider two datasets to compare Info-Style GAN and state-of-the-art disentanglement models: d Sprites (Matthey et al., 2017) and our proposed Isaac3D (See details in Section 4.1). d Sprites is a commonly used dataset in disentanglement learning, with 737,280 images and each of resolution 64x64.
Dataset Splits No The paper does not explicitly provide training, validation, and test dataset splits with specific percentages or counts. It mentions varying the portion of labeled data (η) but not a standard train/validation/test split.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper mentions various models and hyperparameters used (e.g., β-VAE, Factor VAE, Info GAN-CR, Style GAN) and references their original papers, but it does not specify software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow versions, CUDA, specific library versions).
Experiment Setup Yes For VAEs, we set β = 6 for β-VAE, γ = 30 for Factor VAE and β = 8 for β-TCVAE after a grid search over different hyperparameters. ... In experiments, we set ξ = 0.75 in Eq. (5) to be the same with (Berthelot et al., 2019). For the hyperparameters {γG, γE, β, α}, we find that setting γG = β = γ, γE = 0, α = 1 works well across different datasets, where we vary γ {1, 10}.