In this repository, we present new neural architecture search algorithms used to obtain this paper's numerical results:
This paper proposes a new family of algorithms for neural architecture search derived from a new geometrical structure induced by the optimal transport problem on semi-discrete space. This structure was introduced in the paper:
To start an architecture search, follow these steps: Execute run_product.py select the number of threads, max layers, set mutation coefficient (1 is suggested), and number children architectures (4-8). The best architecture is saved every time number of layers changes. To train final architecture, stop and execute post_training: execute run_post_training.py.