ml-experiment-utils

Opinionated utilities for quick ML and deep learning experiments

A personal collection of reusable utilities designed to streamline machine learning and deep learning experimentation workflows. This package provides ready-to-use components for common ML tasks, with built-in best practices and sensible defaults.

Features

Training Loops

• simple_cls_train_v1: Complete training loop for classification tasks with:
  • Automatic device detection (CUDA/CPU)
  • Cosine annealing learning rate scheduling
  • Built-in Weights & Biases integration for experiment tracking
  • Periodic evaluation and logging
  • Progress tracking with tqdm
  • Exponentially weighted moving averages for metrics

Data Utilities

• CIFAR_for_torch: Pre-configured CIFAR-10 dataset with:
  • Standard normalization (channel-wise mean/std)
  • Data augmentation for training (ColorJitter, horizontal flip, rotation)
  • Ready-to-use PyTorch DataLoaders

Helper Classes

• ApproximateSlidingAverage: Efficient exponentially weighted moving average for metric tracking

Installation

pip install ml-experiment-utils

Or with uv:

uv add ml-experiment-utils

Quick Start

Training a Model on CIFAR-10

import torch.nn as nn
from ml_experiment_utils.experiments.data import CIFAR_for_torch
from ml_experiment_utils.experiments.train_loops import simple_cls_train_v1

# Load data
train_loader, test_loader = CIFAR_for_torch(batch_size=128)

# Define your model
model = nn.Sequential(
    nn.Flatten(),
    nn.Linear(3*32*32, 512),
    nn.ReLU(),
    nn.Linear(512, 10)
)

# Train with built-in logging and evaluation
simple_cls_train_v1(
    model=model,
    epochs=10,
    eval_steps=100,
    train_loader=train_loader,
    test_loader=test_loader,
    lr=5e-4,
    name="my-experiment"
)

Using the Data Loaders

from ml_experiment_utils.experiments.data import CIFAR_for_torch

# Get pre-configured CIFAR-10 loaders
train_loader, test_loader = CIFAR_for_torch(
    batch_size=64,
    root="./my_data"
)

# Start training immediately
for images, labels in train_loader:
    # Your training code here
    pass

Requirements

• PyTorch (with torchvision)
• Weights & Biases
• tqdm

Design Philosophy

This package is designed with the following principles:

1. Opinionated but flexible: Sensible defaults that work well for most cases, but configurable when needed
2. Batteries included: Everything you need to start experimenting quickly
3. Experiment tracking first: Built-in W&B integration for reproducibility
4. Minimal boilerplate: Focus on model architecture, not training infrastructure

Use Cases

This package is ideal for:

• Quick prototyping of ML models
• Educational projects and learning
• Baseline implementations for research
• Rapid iteration on model architectures

Contributing

This is primarily a personal utility package, but suggestions and improvements are welcome! Feel free to open an issue or submit a pull request.

License

MIT License - see LICENSE file for details.