andrewssobral / nord

Deep neural architecture research framework

nord

Neural Architecture Search Framework

NORD is a NAS research framework that aims to provide tools in order to make the implementation and comparison of neural architecture design methods fair and straightforward.

Main concepts

• Design Algorithms: Designing neural network specifications

• Descriptors: Making the programmatical generation of nodes and connections more straigth-forward.

• Evaluators: Evaluating the quality of descriptors' networks.

• Environments: Managing the distributed execution of Evaluators.

Main requirements

• PyTorch and Torchvision (https://pytorch.org/)

• NetworkX

• Tensorflow 1.15 for the NASBench-101 benchmark dataset

Examples

• descriptor_example.py Example, usage of NeuralDescriptor and NeuralNet classes
• local_evaluator_example.py, Example usage of LocalEvaluator, which evaluates the given architecture on various datasets.
• nasbench_evaluator_example.py, Example usage of BenchmarkEvaluator,which evaluates the given architecture in NASBench-101.
• reproducible_genetic_algorithm_example_cifar10.py, Example of a reproducible simple genetic algorithm based on DeepNEAT.
• custom_dataset_example.py, Example of evaluating on a custom dataset

Relevant Publications

https://doi.org/10.1016/j.simpa.2020.100042