MAF Benchmark: Preprocessed Datasets for Density Estimation Experiments

Preprocessed versions of seven public datasets used to evaluate the Masked Autoregressive Flow model. George Papamakarios and colleagues from the University of Edinburgh prepared these datasets to ensure reproducibility of their 2017 NeurIPS paper experiments. The collection includes datasets on household power consumption, gas sensor readings, particle physics, and image data.

Use Cases

• Benchmarking density estimation models based on the standard datasets mentioned in the description
• Reproducing experiments for the Masked Autoregressive Flow paper based on the provided preprocessing
• Training generative models on preprocessed tabular and image data like POWER, GAS, and MNIST

Strengths

• Datasets are preprocessed specifically for the experiments described in a peer-reviewed NeurIPS 2017 paper
• Includes seven distinct public datasets covering domains from sensor data to particle physics
• Clear provenance with links to all original data sources

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Last update date is unknown; freshness unverified

Provenance

Source

Original datasets from UCI Machine Learning Repository, Berkeley, Yann LeCun, and others

Collection Method

Preprocessed by the paper authors from public sources for experimental consistency