Autoregressive Quantile Networks for Generative Modeling

Authors: Georg Ostrovski, Will Dabney, Remi Munos

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this work we extend the Pixel CNN model with AIQN and demonstrate results on CIFAR-10 and Image Net using Inception score, FID, non-cherrypicked samples, and inpainting results. We consistently observe that AIQN yields a highly stable algorithm that improves perceptual quality while maintaining a highly diverse distribution. [...] Towards a practical algorithm we base our experimental results on Gated Pixel CNN (van den Oord et al., 2016b), and show that using AIQN significantly im-proves objective performance on CIFAR-10 and Image Net 32x32 in terms of Fr echet Inception Distance (FID) and Inception score, as well as subjective perceptual quality in image samples and inpainting.
Researcher Affiliation Industry 1Deep Mind, London, UK. Correspondence to: Georg Ostrovski <ostrovski@google.com>, Will Dabney <wdabney@google.com>.
Pseudocode No The paper describes the methods using mathematical formulations and textual explanations, but it does not include any figures, blocks, or sections explicitly labeled as "Pseudocode" or "Algorithm".
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We begin by demonstrating Pixel IQN on CIFAR-10 (Krizhevsky & Hinton, 2009). [...] Next, we turn to the small Image Net dataset (Russakovsky et al., 2015), first used for generative modeling in the Pixel RNN work (van den Oord et al., 2016c).
Dataset Splits No The paper mentions evaluating models and hyperparameter search, and refers to an "Appendix for detailed hyperparameters and training procedure." However, the provided text does not include the Appendix, and the main body of the paper does not explicitly specify the exact training, validation, and test dataset splits (e.g., percentages or sample counts).
Hardware Specification Yes Finally, motivated by the very long training time for the large Pixel CNN model (approximately 1 day per 100K training steps, on 16 NVIDIA Tesla P100 GPUs), we also trained smaller 15-layer versions of the models (same as the ones used on CIFAR-10) on the small Image Net dataset. For comparison, these take approximately 12 hours for 100K training steps on a single P100 GPU, or less than 3 hours on 8 P100 GPUs.
Software Dependencies No The paper discusses various models and architectures like Pixel CNN, VAE, and GANs, but it does not specify any software dependencies (e.g., libraries, frameworks, or specific tools) with their version numbers that would be needed for replication.
Experiment Setup No The paper states: "Both models correspond to the 15-layer network variant in (van den Oord et al., 2016b), see Appendix for detailed hyperparameters and training procedure." and "...details can be found in the Appendix." Since the Appendix is not included in the provided text, specific experimental setup details such as hyperparameter values are not present in the main body of the paper.