atalaya

Atalaya

Atalaya is a lightweight toolkit that helps you keep PyTorch experiments organised. It provides:

• a Writer built on top of tensorboardX with handy helpers for scalars, models, CSV exports, and optional integrations with Weights & Biases, Neptune, Comet, and ClearML;
• automatic info.json metadata snapshots (command, Python version, Git details) so you can replay runs later;
• a colour-aware terminal helper for structured CLI output with timestamps;
• a Timer utility to track wall-clock timing for blocks of code or functions;
• optional console log capture so that your scripts remain reproducible.

The writer uses tensorboardX as its event backend, so you get native TensorBoard support while still being able to plug in third-party experiment trackers such as Weights & Biases, Neptune, ClearML, or Comet.

Integration	Status
TensorBoardX	Works (core backend)
Weights & Biases	Works (Writer.with_wandb)
Neptune	Not fully validated yet (Writer.with_neptune)
ClearML	Not fully validated yet (Writer.with_clearml)
Comet ML	Not fully validated yet (Writer.with_comet)

Installation

Install the core package from PyPI:

pip install atalaya

Extras are available if you want to enable third-party integrations:

# Enable Weights & Biases support
pip install atalaya[wandb]

# Install everything (Neptune, Comet, ClearML, matplotlib, seaborn)
pip install atalaya[all]

Quick Start

TensorBoard logging

from atalaya.writer import Writer

writer = Writer(
    name="baseline",
    project="my-awesome-project",
    logdir="logs",
    add_time=True,
    save_as_csv=True,
    output_catcher=True,
    log_git_info=True,
)

# Optional: sync logs with Weights & Biases
writer.with_wandb(group="experiments", entity="my-team")

for epoch in range(10):
    metrics = {"loss": 0.1 * epoch, "accuracy": 0.5 + 0.05 * epoch}
    writer.add_scalars(metrics, global_step=epoch, prefix="train")

    # WandB-style convenience API
    writer.log({"train/loss": metrics["loss"]}, step=epoch)

    # Log entire models to track parameter and gradient histograms
    writer.add_models(
        {"encoder": encoder, "decoder": decoder},
        global_step=epoch,
        log_type="all",  # "parameters", "gradients", or "all"
    )

writer.close()

CSV logging is enabled when save_as_csv=True, and calling output_catcher=True mirrors everything printed to the console into log.txt within the run folder. Every run writes an info.json file with the launch command, Python version, and other metadata in the log directory. When log_git_info is enabled and a Git repository is detected, the same file also includes Git details plus ready-to-run checkout/clone commands so experiments can be reproduced. You can merge your own metadata into the file via the info parameter (for example info={"dataset": "cifar10"}).

Optional integrations

writer.with_wandb(group="experiments", entity="my-team")
writer.with_neptune(entity="my-workspace")

Install the matching extras first (for example poetry add atalaya[wandb] or pip install atalaya[neptune]).

Timing utilities

from atalaya.time_manager import Timer

timer = Timer("training")

with timer:
    run_training_loop()

timer.report(report_type="total_with_stats")  # prints coloured summary to the terminal

Terminal helper

from atalaya.terminal import terminal

# Optional: persist terminal output without a Writer output catcher
terminal.set_log_file("logs/terminal.log")

# Override colours either by name or raw ANSI code
terminal.set_named_color("orange", "\033[38;5;208m")
terminal.set_color("warning", "orange")

terminal.print_info("Loading data...")
terminal.print_warning("Loss is plateauing...", color="orange")
terminal.print_ok("Training finished successfully!")

Messages are timestamped with the process uptime by default. When you enable Writer(..., output_catcher=True) the log file is handled automatically, so calling set_log_file is not required.

Example project

The example/ directory ships with a small PyTorch multi-layer perceptron that demonstrates writer usage end-to-end:

python example/mlp.py

Logs are saved in example/logs/, and the script shows how to combine Writer, Timer, and the terminal helper in a real training loop.

Development and Publishing

This project is managed with Poetry.

# Install dependencies (add --with dev to use Poe tasks)
poetry install --with dev

# Clean previous builds and publish to PyPI
poe publish

# Or run the Poetry command directly
# poetry publish --build

The package metadata (version, dependencies, classifiers) lives in pyproject.toml.