OneShot neural architecture and hardware architecture search
AssigneeGoogle Inc.
InventorsSheng Li, Norman Paul Jouppi, Garrett Axel Andersen, Quoc V. Le, Liqun Cheng, Parthasarathy Ranganathan
Life of this patent
Estimating the term…
Abstract
Aspects of the disclosure are directed to jointly searching machine learning model architectures and hardware architectures in a combined space of models, hardware, and mapping strategies. A search strategy is utilized where all models, hardware, and mappings are evaluated together at once via weight sharing and a supernetwork. A multi-objective reward function is utilized with objectives for quality, performance, power, and area.
Claims17
A method for performing a joint neural architecture and hardware architecture search, the method comprising:
- performing, with one or more processors, a multi-trial search governing multiple OneShot searches for different machine learning models to perform a particular task;
- determining, with the one or more processors, a machine learning model to which a first hardware and a second hardware correspond based on the multi-trial search;
- identifying, with the one or more processors, first parameters for a performance, power, and area (PPA) for the first hardware via an offline-trained machine learning model for predicting PPA;
- determining, with the one or more processors, a first reward based on the first identified parameters;
- identifying, with the one or more processors, second parameters for a PPA for the second hardware via the offline-trained machine learning model, the second parameters being identified based on the first reward;
- determining, with the one or more processors, a second reward based on the second identified parameters;
- and
- selecting, with the one or more processors, hardware corresponding to the machine learning model based on the first reward and the second reward to perform the particular task.
A system comprising:
- one or more processors;
- and
- one or more storage devices coupled to the one or more processors and storing instructions that, when executed by the one or more processors, causes the one or more processors to perform operations for performing a joint neural architecture and hardware architecture search, the operations comprising:
- performing a multi-trial search governing multiple OneShot searches for different machine learning models to perform a particular task;
- determining a machine learning model to which a first hardware and a second hardware correspond based on the multi-trial search;
- identifying first parameters for a performance, power, and area (PPA) for the first hardware via an offline-trained machine learning model for predicting PPA;
- determining a first reward based on the first identified parameters;
- identifying second parameters for a PPA for the second hardware via the offline-trained machine learning model, the second parameters being identified based on the first reward;
- determining a second reward based on the second identified parameters;
- and
- selecting hardware corresponding to the machine learning model based on the first reward and the second reward to perform the particular task.
A non-transitory computer readable medium for storing instructions that, when executed by one or more processors, causes the one or more processors to perform operations for performing a joint neural architecture and hardware architecture search, the operations comprising:
- performing a multi-trial search governing multiple OneShot searches for different machine learning models to perform a particular task;
- determining a machine learning model to which a first hardware and a second hardware correspond based on the multi-trial search;
- identifying first parameters for a performance, power, and area (PPA) for the first hardware via an offline-trained machine learning model for predicting PPA;
- determining a first reward based on the first identified parameters;
- identifying second parameters for a PPA for the second hardware via the offline-trained machine learning model, the second parameters being identified based on the first reward;
- determining a second reward based on the second identified parameters;
- and
- selecting hardware corresponding to the machine learning model based on the first reward and the second reward to perform the particular task.
Dependencies are inferred from the claim text.