Using MO-TPE for neural architecture search

Dependencies

Please install the dependencies via:

pip install -r requirements.txt

Features:

• Neural architecture search as hyperparameter optimization but caring about two objectives: maximizing test set accuracy and minimizing network size.
• Hierachichal search space for convolutional neural nets.
• Multifidelity: implemeted a speedup technique that stops for worst configurations at an early stage (similarly to successive halving)

Original code base made available by paper authors:

Yoshihiko Ozaki

Shuhei Watanabe

Multiobjective tree-structured parzen estimator for computationally expensive optimization problems