neeratyoy / auto-cnn

Nueral Architecture Search and Hyperparameter Optimisation for KMNIST and K49

Auto-configuration of CNNs [Automated Algorithm Design]

The following repo is a sanitized version of the project submitted for the AutoML coursework at University of Freiburg. The src/ folder contains the relevant source code and further details of the code structure.

As on overview:

• The task was to train CNN models for two datasets KMNIST and K49
• With minimal or no manual tuning for the architecture or parameters
• The problem was approached as a Hyper-parameter Optimization (HPO) task
  • With Neural Architecture Search (NAS) as HPO
  • Tuning of training parameters as HPO
• BOHB was used as a tool for HPO
• Overall, transfer learning was leveraged to optimize performance for the datasets

Hardware used:

• All scripts were confined to run one CPU at a time
• Utilizing a single core of Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz

Compute budget:

• A maximum budget of 24 hours on the above specification

Results obtained:

Dataset	Keras simple CNN benchmark	Auto-CNN
KMNIST	95.12%	97.89%
K49	89.25%	94.28%

Note: Results are using simple NAS without skip connections, or specialized architectures (using which can improve results and may increase compute time too)

For KMNIST: A simple BOHB was run for 20 iterations.

For K49: The best returned configuration from KMNIST was chosen. The size of the channels, number of fully connected layers and neurons, batch size was reparameterized and input to BOHB for K49.