pelutils 4.0.0a1

The Swiss army knife of Python projects

pelutils

The Swiss army knife of Python projects

Every project, experiment, or one-off script inevitably ends up reinventing much of the same plumbing: some way to time a loop, saving and loading data to and from disk in a convenient and human-readable manner, a decent logger, parsing a config file, a readable table. pelutils bundles the good versions of these so you can get straight to the actual work. It has no required dependencies beyond the scientific-Python staples, ships type hints (including py.typed), and is easy to start using.

📖 Full documentation: pelutils.readthedocs.io

Highlights

• Logger — easy-to-use, colourful console output, log files with rotation, automatic stacktrace capture, and safe logging from multiple processes.
• Timer & profiler — a Matlab-style tick/tock timer and a near-zero-overhead profiler that prints a readable breakdown of where your time goes.
• UniversalJsonModel — a pydantic.BaseModel that can save any attribute to a human-readable JSON file (numpy arrays, tensors, and other unserialisable types are pickled transparently) and load it straight back.
• JobParser — one parser that unifies command-line arguments and config files, with support for running many jobs from a single config and auto-documenting them.
• unique — a linear-time drop-in for numpy.unique, dramatically faster on large arrays (backed by a small C extension).
• Data-science helpers — a matplotlib Figure context manager with improved default settings over matplotlib, histogram binning, reparametrised scipy distributions, z_score, LaTeX-ready tables, and numpy type aliases.

Installation

pip install pelutils

pelutils supports Python 3.11+. A small subset of functionality can additionally make use of PyTorch, which must be installed separately.

Importing: every feature lives in its own submodule and must be imported from there — e.g. from pelutils.logging import log. Only __version__ is exported at the top level. See the docs for the full API.

Logging

Python's built-in logging is powerful but fiddly to set up, and a bare print gives you no importance levels, no timestamps, and nothing on disk to look at afterwards. This logger hits the sweet spot: one configure call and you get colour-coded, timestamped output to both the console and a log file, with severity levels, log rotation, one-line exception logging, and multiprocessing-safe collection.

from pelutils.logging import log, LogLevels

# Set up the logger by giving it the file to write to
# Omit the path to only print, never write a file
log.configure("train.log")

log.section("Training run")  # Highlighted section header
log(f"Loaded {len(dataset):,} samples")  # Logs at INFO level
for epoch in range(epochs):
    loss = train_one_epoch()
    log.debug(f"Epoch {epoch}: loss {loss:.4f}")  # Logs at DEBUG level - by default, only saved to the log file, but not printed to the console
    if loss > 1e3:
        log.warning("Loss is diverging")  # Logs at WARNING level
log.debug("Final weights", model.state_dict().keys())

# If an exeption occurs in a piece of code, log it with its full (chained) stacktrace, then re-raise
with log.log_errors:
    save_checkpoint(model)

# Temporarily change or silence the log level
with log.level(LogLevels.ERROR):
    log.warning("Suppressed")

# Rotate the log file by time or size
log.configure("train.log", rotation="day")    # or "1 GB", "hour", ...

When using multiprocessing, wrap a worker in with log.collect: so its lines are written together instead of interleaving with other processes. See the logging docs for input helpers, multiple loggers, and more.

Timing and profiling

When you want to know where a script spends its time, cProfile gives you a wall of function-level numbers, and manual time.perf_counter() calls quickly turn into bookkeeping. TickTock sits in between: wrap the sections you care about in named, nestable context managers and print a readable table of totals, hit counts, averages, and each section's share of its parent's time. The per-profile overhead is tiny, so it happily lives inside hot loops and long-running jobs.

from pelutils.ticktock import TT

# Time a single block, Matlab style
TT.tick()
model = train_model(data)
print(f"Training took {TT.tock():.1f} s")

# Profile named sections across a loop, nesting them however you like
for image in images:
    with TT.profile("Process image"):
        with TT.profile("Load"):
            img = load(image)
        with TT.profile("Resize"):
            img = resize(img, (224, 224))
        with TT.profile("Inference"):
            predict(model, img)

# Print a table of hits, total time, and average time for each section
print(TT)

with TT.profile("name", hits=n): records n hits at once, which is handy for very tight loops or for a block that processes n items in parallel. TT.do_at_interval(...) turns the same instance into a throttle for periodic tasks. The default TT is a shared instance; construct your own with TickTock() when you need isolation — most importantly one per thread, as profiling is not thread-safe.

Serialisation

UniversalJsonModel extends pydantic.BaseModel with save/load methods and can serialise attributes that pydantic cannot — numpy arrays, tensors, and arbitrary objects are base64-pickled inline, everything else stays plain, human-readable JSON. Long lists are wrapped to the line-length limit instead of one element per line.

import numpy as np
from pydantic import BaseModel
from pelutils.serialization import UniversalJsonModel
from pelutils.types import FloatArray

class Nested(BaseModel):
    label: str

class Result(UniversalJsonModel):
    accuracy: float
    predictions: FloatArray   # numpy arrays are handled automatically
    meta: Nested

result = Result(
    accuracy=0.97,
    predictions=np.arange(5, dtype=np.float16),
    meta=Nested(label="run-1"),
)

result.save("results/run-1.json")
result = Result.load("results/run-1.json")

Use to_json_dict() / from_json_dict(...) to convert to and from a plain dict without touching the filesystem — useful for nesting inside other structures. The pretty_json helper function is also available on its own. The serialization module also includes JSONL read/write helpers (jsonl_dump, jsonl_load, ...) with largely the same interface as is provided by the built-in json module.

Config and command-line argument parsing

JobParser combines typed command-line options with INI config files. CLI values override config values, which override defaults. Declare RequiredArg for values every job must provide, OptionalArg for values with defaults, and Flag for booleans. Names are --kebab-case on the command line and snake_case attributes on the resulting job.

from pathlib import Path
from pelutils.job_parser import Flag, JobParser, OptionalArg, RequiredArg

parser = JobParser(
    RequiredArg("data-path", help="Training data directory"),
    OptionalArg("learning-rate", default=1e-4, type=float, help="Optimizer learning rate"),
    Flag("fp16", help="Use mixed precision"),
    multiple_jobs=True,
)

for job in parser.parse_jobs():
    print(job.name, job.data_path, job.learning_rate, job.fp16)
    job.write_documentation(Path("runs") / job.name / "arguments.ini")
    # ... run your application with the resolved job values

A single config file can define several named jobs (with a shared [DEFAULT] section), and one CLI override applies to all of them:

python main.py --config-file config.ini --learning-rate 5e-5

For a single job, drop multiple_jobs=True and call parse_job() instead. A config path can target one section directly, e.g. --config-file config.ini:low-lr. See the job parser docs for auto-documentation details.

Fast unique

A linear-time alternative to numpy.unique, significantly faster on large arrays. It also accepts torch tensors and pandas series. The returned elements are unsorted.

import numpy as np
from pelutils.misc import unique

x = np.random.randint(0, 100, size=10_000_000)
values = unique(x)
values, index, inverse, counts = unique(
    x, return_index=True, return_inverse=True, return_counts=True,
)

Data science

Statistics

Common statistical helpers, plus wrappers around scipy distributions reparametrised as in Jim Pitman's Probability (rather than scipy's loc/scale, which are unintuitive for many distributions).

from pelutils.stats import z_score
from pelutils.stats import expon

# 95 % confidence interval half-width for a standard normal (defaults give ~1.96)
half_width = std * z_score()

# One-sided z value for an Exponential(λ=2) at the 1 % significance level
zval = z_score(alpha=0.01, two_sided=False, distribution=expon(lambda_=2))

Plotting

The Figure context manager fixes common matplotlib annoyances — sensible default figure and font sizes, easy styling — and saves and closes the figure for you while restoring rcParams afterwards.

import matplotlib.pyplot as plt
from pelutils.plots import Figure, histogram, normal_binning

with Figure("plot.png", figsize=(20, 10), fontsize=20):
    plt.scatter(x, y, label="Data")
    plt.grid()
    plt.title("Very nice plot")
# Saved to plot.png and closed here

# histogram returns x and y coordinates ready for unpacking
plt.plot(*histogram(data, binning_fn=normal_binning))

Three binning functions are provided — linear_binning, log_binning, and normal_binning (more resolution near the centre of roughly-normal data) — and custom binning functions are supported. See the plotting docs.

Numpy type aliases

Type aliases so you (and your type checker) do not have to track array dtypes by hand.

from pelutils.types import FloatArray, IntArray, BoolArray

def process(features: FloatArray, labels: IntArray, mask: BoolArray): ...

Also included

• pelutils.misc.Table — build aligned text tables which can also be easily export to LaTeX with Table.to_latex().
• pelutils.misc.hardware_info / OS — describe the machine the code runs on.
• pelutils.misc.git_repo_info — the repo and commit the code is executing in.
• Assorted file and dict helpers (reverse_line_iterator, except_keys, ...).
• pelutils.tests — pytest helpers: a UnitTestCollection base class with a managed temp directory, and a restore_argv decorator.

Supported platforms

Precompiled wheels are provided for most common platforms (not 32-bit systems). If no wheel matches, pip builds from source which requires <Python.h> — install it with sudo apt install python3-dev (Ubuntu) or sudo dnf install python3-devel (Fedora).