Examples

An easy way to get started with Auptimizer is to modify the demo code in the Examples folder.

Example	Folder	Purpose
Basic Demo	demo	Tutorial
2D function with different HPOs	2dfunc_diff_opt	Show how to switch between different optimizers
2D function with different resources	2dfunc_diff_res	Show how to switch between different resources
MNIST DNN	hpo_mnist	Show HPO usage for DNN
Tensorflow flags	tf_flags	Show how to integrate with Tensorflow Flags
Tensorflow Iris	tf_iris_diff_opt	Show example on Iris data
NAS integration	cai_nas	Show how to do NAS (uses a publicly available open-source NAS implementation)
Failure Control	job_failure_control	Show job failure control cases

Auto convert

In example 2dfunc_diff_opt, the experiment_auto.json shows how experiment configuration is managed. The user can use:

python -m aup.convert rosenbrock_origin.py experiment_auto.json rosenbrock

to automatically convert the original file to the Auptimizer version. The output file name is defined in experiment_auto.json as script.

Manual modification

We also provide a modified training script in example 2dfunc_diff_opt for users’ reference. In rosenbrock_hpo.py, we show how to manually convert the function for tuning with Auptimizer.

For the end-user’s experiment, simply replacing the rosenbrock() function with their code is enough to use Auptimizer (Need to return the score in that function).

Experiment configurations

Different experiment*.json files in multiple examples illustrate how to specify the configuration for different HPO algorithms. Most of them are identical. To set up a new experiment, please define the corresponding parameter_config in the JSON file.