Smoothing regularization for a generative neural network
A style-based generative network architecture enables scale-specific control of synthesized output data, such as images. During training, the style-based generative neural network (generator neural network) includes a mapping network and a synthesis network. During prediction, the mapping network may be omitted, replicated, or evaluated several times. The synthesis network may be used to generate highly varied, high-quality output data with a wide variety of attributes. For example, when used to generate images of people's faces, the attributes that may vary are age, ethnicity, camera viewpoint, pose, face shape, eyeglasses, colors (eyes, hair, etc.), hair style, lighting, background, etc. Depending on the task, generated output data may include images, audio, video, three-dimensional (3D) objects, text, etc.
1. A computer-implemented method, comprising:
generating output data by a generator neural network based on a set of parameters, wherein the generator neural network comprises one or more layers that each output features to a subsequent layer of the one or more layers;
identifying first features associated with a first layer of the one or more layers and second features associated with a second layer of the one or more layers;
selecting a first modification with respect to the first features;
computing a second modification with respect to the second features, wherein the second modification is consistent with the first modification and computing the second modification comprises;
modifying the first features according to the first modification to yield modified first features;
re-computing the second features based on the modified first features to yield modified second features; and
computing the second modification as the difference between the second features and modified second features;
computing a regularization loss based on the second modification; and
updating the set of parameters to reduce the regularization loss.
2. The computer-implemented method of claim 1 , wherein selecting the first modification comprises selecting each component of the first modification randomly.
3. The computer-implemented method of claim 1 , wherein the first modification is selected from a Gaussian distribution of random values.
4. The computer-implemented method of claim 1 , wherein the generator neural network is a style-based generator neural network.
5. The computer-implemented method of claim 4 , wherein the first features are the output data and the second features are a style signal.
6. The computer-implemented method of claim 4 , wherein the first features are intermediate data and the second features are a style signal.
7. The computer-implemented method of claim 1 , wherein computing the regularization loss comprises computing a magnitude of the second modification.
8. The computer-implemented method of claim 7 , wherein computing the regularization loss further comprises comparing the magnitude against a reference value.
9. The computer-implemented method of claim 8 , wherein the reference value is a constant.
10. The computer-implemented method of claim 8 , further comprising computing the reference value as an average of the magnitude and additional magnitudes over several executions of the generator neural network.
11. The computer-implemented method of claim 8 , wherein updating the set of parameters brings the magnitude closer to the reference value.
12. The computer-implemented method of claim 1 , wherein at least one of the steps of generating, identifying, selecting, computing the second modification, computing the regularization loss, or updating is performed within a cloud computing environment.
13. The computer-implemented method of claim 1 , wherein at least one of the steps of generating, identifying, selecting, computing the second modification, computing the regularization loss, or updating is performed on a server or in a data center to generate an image, and the image is streamed to a user device.
14. The computer-implemented method of claim 1 , wherein at least one of the steps of generating, identifying, selecting, computing the second modification, computing the regularization loss, or updating is performed to generate an image used for training, testing, or certifying a neural network employed in a machine, robot, or autonomous vehicle.
15. A computer-implemented method, comprising:
generating output data by a generator neural network based on a set of parameters, wherein the generator neural network comprises one or more layers that each output features to a subsequent layer of the one or more layers;
identifying first features associated with a first layer of the one or more layers and second features associated with a second layer of the one or more layers, wherein generating the output data by the generator neural network comprises computing the second features before computing the first features;
selecting a first modification with respect to the first features;
computing a second modification with respect to the second features, wherein the second modification is consistent with the first modification;
computing a regularization loss based on the second modification; and
updating the set of parameters to reduce the regularization loss.
16. The computer-implemented method of claim 15 , wherein computing the second modification comprises:
computing an inner product between the first features and the first modification; and
differentiating the inner product with respect to the second modification.
17. A system, comprising:
a processor configured to implement a generator neural network comprising one or more layers that each output features to a subsequent layer of the one or more layers, wherein the generator neural network is configured to:
generate output data based on a set of parameters;
identify first features associated with a first layer of the one or more layers and second features associated with a second layer of the one or more layers;
select a first modification with respect to the first features;
compute a second modification with respect to the second features, wherein the second modification is consistent with the first modification and computing the second modification comprises:
modifying the first features according to the first modification to yield modified first features;
re-computing the second features based on the modified first features to yield modified second features; and
computing the second modification as the difference between the second features and modified second features;
compute a regularization loss based on the second modification; and
update the set of parameters to reduce the regularization loss.
18. The system of claim 17 , wherein selecting the first modification comprises selecting each component of the first modification randomly.
19. A non-transitory computer-readable media storing computer instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of:
generating output data by a generator neural network based on a set of parameters, wherein the generator neural network comprises one or more layers that each output features to a subsequent layer of the one or more layers;
identifying first features associated with a first layer of the one or more layers and second features associated with a second layer of the one or more layers;
selecting a first modification with respect to the first features;
computing a second modification with respect to the second features, wherein the second modification is consistent with the first modification and computing the second modification comprises:
modifying the first features according to the first modification to yield modified first features;
re-computing the second features based on the modified first features to yield modified second features; and
computing the second modification as the difference between the second features and modified second features;
computing a regularization loss based on the second modification; and
updating the set of parameters to reduce the regularization loss.
20. The non-transitory computer-readable media of claim 19 , wherein selecting the first modification comprises selecting each component of the first modification randomly.