NOTE: This reader is now included in NumPy and is used for `np.loadtxt`. The code in this repository is not up to date and the NumPy version should be used. There are known bugs or subtle differences only fixed in NumPy.

Read text files (e.g. CSV or other delimited files) into a NumPy array.

npreadtext has been tested with NumPy v1.18 and higher and can be installed using:

python -m pip install numpy
python -m pip install git+git:https://github.com/BIDS-numpy/npreadtext

To enable the C-accelerated version of np.loadtxt, monkey-patch NumPy:

import numpy as np
from npreadtext import monkeypatch_numpy

This replaces np.loadtxt with npreadtext._loadtxt.

For more detailed information on installation, testing, and benchmarking - see below.

Requires NumPy:

pip install -r requirements.txt

To run the test and benchmarking suites, you will need some additional tools:

pip install -r dev_requirements.txt

Build and install w/ pip: pip install -e .. The --verbose flag is useful for seing build logs: pip install -e . --verbose. Full (syntax-highlighted) build log also via python setup.py build_ext -i.

There are three sets of tests:

• npreadtxt test suite:

  pytest .
  

• Compatibility with np.loadtxt:

  python compat/check_loadtxt_compat.py -t numpy.lib.tests.test_io::TestLoadTxt
  

The following is a quick-and-dirty procedure for evaluating the performance of npreadtext with the numpy benchmark suite. TODO: figure out how to get configure asv to do this comparison directly. The pain point was getting npreadtext installed in the virtual environments that asv creates. This is a hacky procedure to work around these complications by running everything in the same virtualenv and falling back on basic utils.

• Create new (empty) virtualenv

• In numpy repo:

  • pip install -r test_requirements.txt
  • pip install -e .
  • pip install asv virtualenv

• In this repo:

  • pip install -e .

• Back in numpy repo, create a branch (asv works best with committed changes):

  • git checkout -b monkeypatch-npreadtxt

  • Modify the numpy/__init__.py to monkeypatch _loadtxt into numpy in place of np.loadtxt. For example, delete the original loadtxt from __init__.py and modify the __getattr__ to return _loadtxt:

    del loadtxt
    def __getattr__(attr):
        if attr == "loadtxt":
            sys.path.append("/path/to/npreadtext/")
            from npreadtext import _loadtxt
            return _loadtxt
        ...
    

  • Commit the changes

In the numpy repo, checkout the branch you want to compare against (presumably main):

• git checkout main
• python runtests.py --bench-compare monkeypatch-npreadtxt bench_io

There is also a script bench/bench.py to facilitate basic performance comparisons with other text loaders such as pd.read_csv. The script uses the IPython %timeit magic so should be run with ipython, e.g.

ipython -i bench/bench.py

By default, pandas.read_csv uses an approximate method for parsing floating point numbers. In practice, this results in faster float parsing at the expense of faithful full-precision reproduction of floating point values on reading/writing. Full-precision float parsing can be selected using the float_precision="round-trip" option of pandas.read_csv.

See also:

• https://pandas.pydata.org/docs/user_guide/io.html#specifying-method-for-floating-point-conversion
• pandas-dev/pandas#17154