ultimate-utils 0.10.6

Brando's Ultimate Utils for Science, Machine Learning, and AI

ultimate-utils (uutils)

A comprehensive Python utility library for machine learning research, spanning PyTorch training infrastructure, statistical analysis, meta-learning, LLM evaluation, and experiment management.

Built and maintained by Brando Miranda (Stanford, MIT, UIUC).

pip install ultimate-utils

import uutils
uutils.hello()

What's Inside

Module	Description
uutils	Core utilities — file I/O, argument parsing, serialization (dill/pickle/JSON), git helpers, seed management, progress bars
uutils.torch_uu	PyTorch training loops, distributed training, checkpointing, optimizers, learning rate schedulers
uutils.torch_uu.models	Model utilities and HuggingFace integrations
uutils.torch_uu.dataloaders	Dataloaders for meta-learning (miniImageNet, CIFAR-FS), standard vision, and multi-dataset sampling
uutils.torch_uu.metrics	CCA/PWCCA/DCCA similarity, model complexity, task diversity (Task2Vec), accuracy with confidence intervals
uutils.torch_uu.meta_learners	Meta-learning algorithms (MAML, Prototypical Networks, etc.)
uutils.stats_uu	Hypothesis testing, p-values, effect sizes (Cohen's d), power analysis, confidence intervals, ANOVA, regression
uutils.plot	Plotting with error bands, heatmaps, bar charts, LaTeX table export
uutils.hf_uu	HuggingFace training utilities — full fine-tuning, QLoRA/Unsloth, causal and seq2seq LM training
uutils.evals	LLM evaluation — math benchmarks (MATH, Putnam, OlympiadBench), API inference (Claude, OpenAI, vLLM), answer extraction
uutils.dspy_uu	DSPy-based synthetic data generation for in-context learning and fine-tuning
uutils.jax_uu	JAX multi-head attention, layer norm, flash attention implementations
uutils.numpy_uu	Statistical moments, confidence intervals, matrix utilities
uutils.logging_uu	Weights & Biases integration — setup, logging, sweeps, model watching
uutils.emailing	SMTP email notifications with attachments

Installation

Note: PyTorch must be installed separately (with CUDA if you need GPU support).

Development install (recommended)

conda create -n uutils python=3.11 -y
conda activate uutils
pip install -e ~/ultimate-utils

pip install from PyPI

pip install ultimate-utils

Verify installation

python -c "import uutils; uutils.hello()"
python -c "import uutils; uutils.torch_uu.gpu_test_torch_any_device()"

Quick Examples

Seed everything for reproducibility

from uutils import seed_everything
seed_everything(42)

Save and load with dill

from uutils import save_with_dill, load_with_dill
save_with_dill(my_object, '~/data/my_object.pkl')
obj = load_with_dill('~/data/my_object.pkl')

Plot with error bands

from uutils.plot import plot_with_error_bands
plot_with_error_bands(x, y, yerr, xlabel='Steps', ylabel='Loss', title='Training Loss')

Statistical testing with effect size

from uutils.stats_uu.effect_size import stat_test_with_effect_size_as_emphasis
stat_test_with_effect_size_as_emphasis(group1_data, group2_data)

W&B logging

from uutils.logging_uu.wandb_logging.common import setup_wandb, log_2_wandb
setup_wandb(args)
log_2_wandb(metrics_dict, step=step)

Publishing to PyPI

# Bump version in setup.py, then:
cd ~/ultimate-utils && bash scripts/publish_to_pypi.sh

Citation

If you use ultimate-utils in your research, please cite:

@software{miranda2024uutils,
  author       = {Brando Miranda},
  title        = {ultimate-utils: A Comprehensive Utility Library for Machine Learning Research},
  year         = {2024},
  publisher    = {PyPI},
  url          = {https://github.com/brando90/ultimate-utils},
  note         = {Available at \url{https://pypi.org/project/ultimate-utils/}}
}

You can also find the author's publications on Google Scholar.

Related Publications

This library has supported research in the following publications (among others):

• Beyond Scale: the Diversity Coefficient as a Data Quality Metric Demonstrates LLMs are Not Yet More Human than Human
• Does MAML Only Work via Feature Re-use? A Data Set Centric Perspective

License

Apache-2.0