Model Search is domain independent and designed to find the right architecture for a given data set or problem. It should keep the coding time as well as the effort and the computing resources as low as possible. Model Search is based on the TensorFlow machine learning framework and can either run on a single machine or in a distributed setting.
A look under the hood
Google’s open source platform consists of several trainer processes, a search algorithm, a transfer learning algorithm and a database for storing the evaluated models. It performs the training and evaluation experiments for AI models in an adaptive and asynchronous manner. This means that all trainer processes share the results from their experiments with other trainers while they conduct each experiment independently. A cycle begins with the search algorithm reviewing all completed experiments, which then decides what to try next. He then calls up the mutation using one of the best architectures found so far and assigns the resulting model to a trainer again.
After a model search, users can compare the models found during the search. In addition, they have the option of creating their own search space in order to adapt the architectural elements in their models. More information can be found in the post on the Google AI blog.