Generative neural network distillation
A compact generative neural network can be distilled from a teacher generative neural network using a training network. The compact network can be trained on the input data and output data of the teacher network. The training network train the student network using a discrimination layer and one or more types of losses, such as perception loss and adversarial loss.
1. A method comprising:
identifying, using a plurality of first neural networks, an object in an image, wherein each of the plurality of first neural networks is trained to recognize one object of a plurality of objects;
selecting a second neural network trained to apply an effect to images that comprise the identified object; and
generating a result image by using the second neural network to apply the effect to the image, wherein the plurality of second neural networks are trained on input data and output data generated by a teacher generative neural network using adversarial loss.
2. The method of claim 1 , wherein the method is performed on a mobile device, and wherein the method further comprises:
accessing input from a user of the mobile device to apply the effect to the image.
3. The method of claim 1 wherein the second neural network is one of a plurality of second neural networks, and wherein one of the plurality of second neural networks is trained to apply the effect for each object of the plurality of objects.
4. The method of claim 3 wherein the method is performed on a mobile device, and wherein the plurality of second neural networks are stored on the mobile device.
5. The method of claim 1 wherein the method is performed on a mobile device.
6. The method of claim 1 :
wherein a second neural network is trained for each object of the plurality of objects.
7. The method of claim 6 further comprising:
identifying the input data and the output data of a teacher generative neural network, the output data generated by modifying the input data using the teacher generative neural network.
8. The method of claim 1 wherein the plurality of second neural networks are trained using a discriminative neural network that evaluates data output by the plurality of second neural networks compared with target data.
9. The method of claim 1 wherein effect is at least one of a painting style transfer, an aging style transfer, a wrinkle remover, a youth style transfer, or an aging style transfer.
10. The method of claim 1 wherein the object is at least one of teeth, a car, or an apple.
11. The method of claim 1 further comprising:
causing the result image to be published as an ephemeral message on a social network site.
12. The method of claim 1 wherein the second neural network is trained based on default losses for the effect and an additional loss specific to the object.
13. The method of claim 12 wherein the default losses comprise: a perception loss, an adversarial loss, and a high-frequency loss.
14. A system comprising:
one or more processors of a machine; and
one or more memories storing instructions that, when executed by the one or more processors, cause the machine to perform operations comprising:
identifying, using a plurality of first neural networks, an object in an image, wherein each of the plurality of first neural networks is trained to recognize one object of a plurality of objects;
selecting a second neural network trained to apply an effect to images that comprise the identified object; and
generating a result image by using the second neural network to apply the effect to the image, wherein the plurality of second neural networks are trained on input data and output data generated by a teacher generative neural network using adversarial loss.
15. The system of claim 14 , wherein the system is a mobile device, and wherein the operations further comprise:
accessing input from a user of the mobile device to apply the effect to the image.
16. The system of claim 14 wherein the second neural network is one of a plurality of second neural networks, and wherein one of the plurality of second neural networks is trained to apply the effect for each object of the plurality of objects.
17. The system of claim 16 wherein the operations are performed on a mobile device, and wherein the plurality of second neural networks are stored on the mobile device.
18. A non-transitory machine-readable storage medium embodying instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
identifying, using a plurality of first neural networks, an object in an image, wherein each of the plurality of first neural networks is trained to recognize one object of a plurality of objects;
selecting a second neural network trained to apply an effect to images that comprise the identified object; and
generating a result image by using the second neural network to apply the effect to the image, wherein the plurality of second neural networks are trained on input data and output data generated by a teacher generative neural network using adversarial loss.
19. The non-transitory machine-readable storage medium of claim 18 , wherein the one or more processors are part of a mobile device, and wherein the operations further comprise:
accessing input from a user of the mobile device to apply the effect to the image.
20. The non-transitory machine-readable storage medium of claim 18 wherein the second neural network is one of a plurality of second neural networks, and wherein one of the plurality of second neural networks is trained to apply the effect for each object of the plurality of objects.