keras-tuner-tutorial

Hands on tutorial for keras-tuner

This repo aims at introducing hyperparameter tuning through the Keras Tuner library. It provides a comparison of its different tuners, applied to computer vision through the CIFAR10 dataset.

This is work in progress, all feedback is welcomed.

Install the project

• Clone the repo
• Create a virtualenv and activate it:

virtualenv -p python3 venv
source venv/bin/activate

• Install the requirements:

pip install requirements.txt

Results

Tasks duration was measured on an RTX 2080 GPU

Tuner	Search time	Best accuracy (%)
Worst Baseline	20min	63.1
Default Baseline	20min	74.5
Random Search	10h 59min	76.8
Hyperband	10h 0min	75.1

Here, the worst baseline is the worst accuracy obtained by a set of hyperparameters during random search. The default baseline is obtained by setting all hyperparameters to their default value.

Run the baseline

python baseline.py

Run the comparison

Available tuners :

• Random Search
• Hyperband

python tuner_comparison.py