ttv - create train, test, validation sets
ttv is a command line tool for splitting large files up into chunks suitable for train/test/validation splits for machine learning. It arose from the need to split files that were too large to fit into memory to split, and the desire to do it in a clean way.
ttv requires Rust 2021.
Installation
Build using cargo build --release to get a binary at ./target/release/ttv. Copy this into your path to use it.
Usage
Run ttv --help to get help, or infer what you can from one of these examples:
# Split CSV file into two sets of a fixed number of rows
$ ttv split data.csv --rows=train=9000 --rows=test=1000 --uncompressed-input
# Accepts gzipped data (no flag required). Shorthand argument version. As many splits as you like!
$ ttv split data.csv.gz --rows=train=65000,validation=15000,test=15000
# Alternatively, specify proportion-based splits. -u is shorthand for --uncompressed-input
$ ttv split data.csv --props=train=0.8,test=0.2 -u
# When using proportions, include the total rows to get a progress bar
$ ttv split data.csv --props=train=0.8,test=0.2 --total-rows=1234 -u
# Accepts data from stdin, compressed or not (must give a filename)
$ cat data.csv | ttv split --rows=test=10000,train=90000 --output-prefix data -u
$ cat data.csv.gz | ttv split --rows=test=10000,train=90000 --output-prefix data
# Using pigz for faster decompression
$ pigz -dc data.csv.gz | ttv split --prop=test=0.1,train=0.9 --chunk-size 5000 --output-prefix data -u
# Split outputs into chunks for faster writing/reading later
$ ttv split data.csv.gz --rows=test=100000,train=900000 --chunk-size 5000
# Write outputs uncompressed
$ ttv split data.csv.gz --prop=test=0.5,train=0.5 --uncompressed-output
# Reproducible splits using seed
$ ttv split data.csv.gz --prop=test=0.5,train=0.5 --chunk-size 1000 --seed 5330
Development
You'll need a recent version of the Rust nightly toolchain and Cargo. Then just hack away as normal.