The AI Models Architecture Search provides automatic designing of deep network architectures to facilitate the development of AI models for developers. Given the specified Quality of Service (QoS) requirements, the algorithm searches for the best network configuration that complies with such requirements. Given the available labelled samples, the Pareto-Optimal Progressive Neural Architecture Search (POPNAS) efficiently identifies the set of candidate models from the Pareto front of the trade-off between time and accuracy.
The open source is provided through the Apache 2.0 licence.
The high-level architecture of the Pareto-Optimal Progressive Neural Architecture Search is depicted in the following figure. After model expansion, the controller evaluates the accuracy of children architectures, while another predictor, named the regressor (obtained from the performance model component), predicts the required time. The output is the most accurate cell structure among the ones with the lowest training time, where a cell is a composition of multiple blocks, i.e., the binary operations searched by the algorithm. At each iteration, the Pareto front is built and only the best K models are selected for the next expansion step. The current implementation is based on Python and relies on standard Python libraries.