Generative data augmentation with task loss guided fine-tuning
AssigneeQUALCOMM Incorporated
InventorsDavide Abati, Amirhossein Habibian, Auke Joris Wiggers, Jens Petersen
Life of this patent
Estimating the term…
Abstract
A method includes generating a synthetic dataset with a generative model. The method also includes tuning the generative model based on feedback from a task network that receives the synthetic dataset as input. The task network may perform image recognition. The synthetic dataset may be generated based on a set of classes and labels of the classes. The method may iteratively generate the synthetic dataset and tune the generative model, based on feedback from the task network.
Claims16
An apparatus, comprising:
- at least one memory storing instructions;
- and
- at least one processor coupled to the at least one memory and configured to instruct the apparatus to:
- generate a synthetic dataset with a generative model;
- obtain a prediction, with a task network, based on the synthetic dataset;
- calculate a loss of the task network based on the prediction from the synthetic dataset;
- back propagate the loss of the task network to the generative model instead of a task model;
- tune the generative model by reducing the loss of the task network to obtain a tuned generative model;
- and
- generate an updated synthetic dataset with the tuned generative model.
A method, comprising:
- generating a synthetic dataset with a generative model based on a set of real images;
- obtaining a first prediction, with a task network, based on the synthetic dataset;
- calculating a first loss of the task network based on the first prediction from the synthetic dataset;
- updating the task network based on the first loss of the task network to obtain an updated task network;
- calculating a second loss of the updated task network based on a second prediction of the updated task network and the set of real images;
- back propagating the second loss of the updated task network to the generative model;
- tuning the generative model by reducing the loss of the updated task network to obtain a tuned generative model;
- and
- generating an updated synthetic dataset with the tuned generative model.
A non-transitory computer-readable medium having program code recorded thereon, the program code executed by a processor and comprising:
- program code to generate a synthetic dataset with a generative model;
- program code to obtain a prediction, with a task network, based on the synthetic dataset;
- program code to calculate a loss of the task network based on the prediction from the synthetic dataset;
- program code to back propagate the loss of the task network to the generative model instead of a task model;
- program code to tune the generative model by reducing the loss of the task network to obtain a tuned generative model;
- and
- program code to generate an updated synthetic dataset with the tuned generative model.
An apparatus, comprising:
- means for generating a synthetic dataset with a generative model based on a set of real images;
- means for obtaining a first prediction, with the task network, based on the synthetic dataset;
- means for calculating a first loss of the task network based on the first prediction from the synthetic dataset;
- means for updating the task network based on the first loss of the task network to obtain an updated task network;
- means for calculating a second loss of the updated task network based on a second prediction of the updated task network and the set of real images;
- means for back propagating the second loss of the updated task network to the generative model;
- means for tuning the generative model by reducing the loss of the updated task network to obtain a tuned generative model;
- and
- means for generating an updated synthetic dataset with the tuned generative model.
Dependencies are inferred from the claim text.