subsample is a command-line tool for sampling data from a large,
newline-separated dataset (typically a CSV-like file).
subsample is distributed with
pip. Once you've installed
> pip install subsample
subsample will be installed into your Python environment.
subsample requires one argument, the input file. If the input file
-, data will be read from standard input (in this case, only
the reservoir and approximate algorithms can be used).
To take a sample of size 1000 from the file
subsample as follows:
> subsample -n 1000 big_data.csv
This will print 1000 random lines from the file to the terminal.
Usually we want to save the sample to another file instead.
subsample doesn't have file output built-in; instead it relies
on the output redirection features of your terminal. To save
big_data_sample.csv, run the following command:
> subsample -n 1000 big_data.csv > big_data_sample.csv
Sampling from STDIN
To use standard input as the source, use - as the filename, eg:
> subsample -n 1000 < big_data.csv > big_data_sample.csv
Note that only reservoir sampling supports stdin because the other sampling algorithms require a seekable input stream.
CSV files often have a header row with the column names. You can pass
-r flag to
subsample to preserve the header row:
> subsample -n 1000 big_data.csv -r > big_data_sample.csv
Rarely, you may need to sample from a file with a header spanning
multiple rows. The
-r argument takes an optional number of
rows to preserve as a header:
> subsample -n 1000 -r 3 data_with_header.csv > sample_with_header.csv
Note that if the
-r argument is directly before the input filename,
it must have an argument or else it will try to interpret the input
filename as the number of header rows and fail. Putting the
after the input filename will avoid this.
The output of
subsample is random and depend on the computer's random
state. Sometimes you may want to take a sample in a way that can be
reproduced. You can pass a random seed to
subsample with the
to accomplish this:
> subsample -s 45906345 data_file.csv > reproducable_sample.csv
subsample implements three sampling algorithms, each with their own strengths
|fixed sample size||compatible||not compatible||compatible|
|fractional sample size||not compatible||compatible||compatible|
For space complexity,
ss is the number of records in the sample and
rs is the maximum size of a record.
Sample order is the order of the records returned. Only reservoir sampling gives results in random order; approximate and two-pass return results in the same order as the source data.
Reservoir sampling (Random Sampling with a Reservoir (Vitter 85)) is a method of sampling from a stream of unknown size where the sample size is fixed in advance. It is a one-pass algorithm and uses space proportional to the amount of data in the sample.
Reservoir sampling is the default algorithm used by
subsample. For consistency,
it can also be invoked with the argument
When using reservoir sampling, the sample size must be fixed rather than fractional.
> subsample --reservoir -n 1000 big_data.csv > sample_data.csv
Approximate sampling simply includes each row in the sample with a probability
given as the sample proportion. It is a stateless algorithm with minimal space
requirements. Samples will have on average a size of
fraction * population_size,
but it will vary between each invocation. Because of this, approximate sampling
is only useful when the sample size does not have to be exact (hence the name).
> subsample --approximate -f 0.15 my_data.csv > my_sample.csv
Equivalently, supply a percentage instead of a fraction by switching the
-f to a
> subsample --approximate -p 15 my_data.csv > my_sample.csv
As the name implies, two-pass sampling uses two passes: the first is to count the
number of records (ie. the population size) and the second is to emit the records
which are part of the sample. Because of this it is not compatible with
as an input.
> subsample --two-pass -n 1000 my_data.csv > my_sample.csv
Two-pass sampling also accepts the sample size as a fraction or percent:
> subsample --two-pass -p 15 my_data.csv > my_sample.csv
A simple GNU Make-driven testing script is included. Run
make test from
subsample's base directory after installing to run some regression tests.
Due to the randomness inherent to random sampling, testing is limited to checking that the output is the same when the random seed is unchanged. This serves mainly to find new bugs introduced by changes in the future and does not imply that the code itself is correct (in the sense that the sample is truly random).