POPNASv3: a Pareto-Optimal Neural Architecture Search Solution for Image and Time Series Classification

POPNASv3 is a neural architecture search algorithm that employs a sequential model-based optimization search strategy to find optimal architectures inside vast search spaces. Our method can find promising architectures for image and time series classification tasks. The algorithm requires a dataset and a configuration file as inputs, returning the best architecture found as output, provided as an ONNX file that contains both the model and the weights trained to convergence.

The popnas-3.0.0 folder contains the source code of the algorithm. The experiments folder contains some output files for each experiment presented in the article, i.e., the ONNX of the final model, a pdf image representing the direct acyclic graph of the cell used in the final model, the results of all networks sampled during the search (after proxy training), and the json configuration used as input for the search procedure.