automl

README

AutoML Hyperparameter Search

Demonstrates automated hyperparameter optimization using Bayesian search with early stopping via the training/automl package.

What this example shows

1. Search space definition -- defining hyperparameters with ranges and log-scale options
2. Bayesian optimization -- using automl.NewBayesianOptimizer as the search strategy
3. Coordinator loop -- running automl.NewCoordinator with a worker, max trials, and early stopping
4. Result reporting -- extracting the best trial configuration

Build

go build -o automl ./examples/automl/

Run

./automl

Expected output

=== AutoML Hyperparameter Search Example ===

Search space:
  lr: [0.0001, 0.1000] (log scale)
  hidden_dim: [32.0000, 256.0000]
  num_layers: [1.0000, 4.0000]
  dropout: [0.0000, 0.5000]

Max trials: 20, Early stop patience: 5

--- Running trials ---
  Trial 1: lr=0.0234 hidden=187 layers=3 dropout=0.42 -> score=0.7043
  Trial 2: lr=0.0012 hidden=95 layers=2 dropout=0.15 -> score=0.8921
  ...

=== Search Complete ===
Best trial: #N
Best score: 0.XXXX
Best hyperparameters:
  lr = ...
  hidden_dim = ...
  num_layers = ...
  dropout = ...

Total evaluations: N

The Bayesian optimizer starts with random exploration, then uses Expected Improvement to focus on promising regions of the search space. Early stopping halts the search if no improvement is found for 5 consecutive trials.

Key APIs

Type	Package	Purpose
automl.HParam	training/automl/	Hyperparameter definition (name, min, max, log scale)
automl.BayesianOptimizer	training/automl/	Bayesian search strategy with Expected Improvement
automl.Coordinator	training/automl/	Orchestrates trial dispatch and early stopping
automl.Worker	training/automl/	Interface for trial evaluation (implement RunTrial)

Production usage

Replace syntheticWorker with a worker that:

1. Builds a model from the suggested hyperparameters
2. Trains for a few epochs on your dataset
3. Returns the validation metric as automl.Metric{Score: ...}

For architecture search across tabular/time-series models, use automl.AutoML() which searches over MLP, FT-Transformer, TabNet, SAINT, TabResNet, TFT, N-BEATS, and PatchTST simultaneously.

Documentation

Overview

Command automl demonstrates using the AutoML coordinator to search over hyperparameter configurations with Bayesian optimization and early stopping.

This example uses a synthetic scoring function. In production, replace the worker with one that trains and evaluates a real model.

Usage:

go build -o automl ./examples/automl/
./automl