Perceptual Generative Autoencoders

Authors: Zijun Zhang, Ruixiang Zhang, Zongpeng Li, Yoshua Bengio, Liam Paull

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate the performance of LPGA and VPGA on three image datasets, MNIST (Le Cun et al., 1998), CIFAR-10 (Krizhevsky & Hinton, 2009), and Celeb A (Liu et al., 2015). For each model and each dataset, we take 5,000 generated samples to compute the FID score. The results (with standard errors of 3 or more runs) are summarized in Table. 1.
Researcher Affiliation Academia 1University of Calgary, Canada 2MILA, Universit e de Montr eal, Canada 3Wuhan University, China.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/zj10/PGA.
Open Datasets Yes In this section, we evaluate the performance of LPGA and VPGA on three image datasets, MNIST (Le Cun et al., 1998), CIFAR-10 (Krizhevsky & Hinton, 2009), and Celeb A (Liu et al., 2015).
Dataset Splits No The paper mentions tuning hyperparameters heuristically but does not provide specific train/validation/test dataset split percentages, counts, or explicit references to standard splits used for validation.
Hardware Specification No All experiments are performed on a single GPU. This statement is too general and does not provide specific model numbers or types of GPU, CPU, or other hardware components.
Software Dependencies No The paper does not explicitly list specific software dependencies with version numbers (e.g., Python, PyTorch versions).
Experiment Setup Yes SGD with a momentum of 0.9 is used to train all models. For LPGA, γ (Eq. (9)) tends to vary in a small range for different datasets (e.g., 1.5e 2 for MNIST and CIFAR-10, and 1e 2 for Celeb A). For VPGA, η (Eq. (13)) can vary widely (e.g., 2e 2 for MNIST, 3e 2 for CIFAR-10, and 2e 3 for Celeb A), and thus is slightly more difficult to tune.