This package uses adaptive evolutionary algorithms in order to train, evaluate, and evolve Convolutional Neural Networks. It includes a simple interface with TensorFlow and allows for parallel model training across multiple GPUs.
The provided code uses the CIFAR 10 dataset as input. The can be downloaded here. The dataset should be saved in the 'datasets/cifar10' directory. Other datasets can be adapted to use the DataSet class provided in dataset.py.
Tournament defautls can be set in defaults.py.
- Python 3
- Numpy
- TensorFlow
- namedlist
- Pickle
For full list of results please see the Report. Example tournament results showing mutations applied at each generation and resulting CNN performance.
- Anastasiya Lazareva - Initial work - alazareva