README.md

Data Handling in Benchmarks

Data handling steps:

1. Load data:
   • If not cached: download/generate dataset and put it in raw and/or usual cache
   • If cached: load from cached files
2. Split data into subsets if requested
3. Convert to requested form (data type, format, order, etc.)

There are two levels of caching with corresponding directories: raw cache for files downloaded from external sources, and just cache for files applicable for fast-loading in benchmarks.

Each dataset has few associated files in usual cache: data component files (x, y, weights, etc.) and JSON file with dataset properties (number of classes, clusters, default split arguments). For example:

data_cache/
...
├── mnist.json
├── mnist_x.parq
├── mnist_y.npz
...

Cached file formats:

Format	File extension	Associated Python types
Parquet	.parq	pandas.DataFrame
Numpy uncompressed binary dense data	.npz	numpy.ndarray, pandas.Series
Numpy uncompressed binary CSR data	.csr.npz	scipy.sparse.csr_matrix

Existing data sources:

• Synthetic data from sklearn
• OpenML datasets
• Custom loaders for named datasets

---

Documentation tree