Optimizing supervised generative adversarial networks via latent space regularizations
A method for image generation based on a Generative Adversarial Network (GAN) including a generator, a discriminator, and an encoder, wherein outputs of the generator are mapped, by the encoder, to a latent space adaptable to manipulate at least one characteristics of images generated by the GAN, the method including generating, by the encoder, a first encoding E(Y) of a target image Y and a second encoding E(G(Z)) of a generated image G(Z) corresponding to the target image Y, wherein the first and second encodings E(Y) and E(G(Z)) map Y and G(Z) to the latent space having a lower dimension than dimensionality of Y and G(Z), wherein the encoder is trained to minimize the differences between the first and second encodings E(Y) and E(G(Z)), and the generator is trained by using the first and second encodings E(Y) and E(G(Z)) as part of a loss function.
1. An apparatus comprising:
at least one processor; and
at least one memory, wherein the at least one memory comprises instructions which, when executed by the at least one processor, cause the at least one processor to implement a Generative Adversarial Network (GAN), the GAN comprising a generator, a discriminator, and an encoder, wherein outputs of the generator are mapped, by the encoder, to a latent space adaptable to manipulate at least one characteristics of images generated by the GAN, and to train the GAN by:
generating, by the encoder, a first encoding E(Y) of a target image Y and a second encoding E(G(Z)) of a generated image G(Z) corresponding to the target image Y, wherein the generated image G(Z) corresponding to the target image Y is generated by the generator, the first encoding E(Y) and the second encoding E(G(Z)) map the target image Y and the generated image G(Z) to the latent space having a lower dimension than dimensionality of the target image Y and the generated image G(Z), and a first loss function between the first encoding E(Y) and the second encoding E(G(Z)) is minimized for the encoder; and
training the generator using the first encoding E(Y) and the second encoding E(G(Z)) as part of a second loss function for the generator.
2. The apparatus of claim 1 , wherein a same set of weights are used for first encoding E(Y) of the target image Y and the second encoding E(G(Z)) of the generated image G(Z) corresponding to the target image Y, and the same set of weights are updated after one or more iterations of training the GAN.
3. The apparatus of claim 1 , wherein the encoder comprises a machine learning model.
4. The apparatus of claim 1 , wherein the encoder comprises a convolutional neural network.
5. The apparatus of claim 1 , wherein the discriminator, the generator, and the encoder are trained, in sequential order, in each iteration of training the GAN.
6. The apparatus of claim 1 , wherein for each iteration of training the GAN, the generator and the encoder are updated simultaneously after the discriminator has been updated.
7. The apparatus of claim 1 , wherein the encoder is adapted from at least one of a first few layers of a VGG model, or one or more upscaling layers following the first few layers of the VGG model.
8. A method for image generation based on a Generative Adversarial Network (GAN), the GAN comprising a generator, a discriminator, and an encoder, wherein outputs of the generator are mapped, by the encoder, to a latent space adaptable to manipulate at least one characteristics of images generated by the GAN, the method comprising:
generating, by the encoder, a first encoding E(Y) of a target image Y and a second encoding E(G(Z)) of a generated image G(Z) corresponding to the target image Y, wherein the generated image G(Z) corresponding to the target image Y is generated by the generator, the first encoding E(Y) and the second encoding E(G(Z)) map the target image Y and the generated image G(Z) to the latent space having a lower dimension than dimensionality of the target image Y and the generated image G(Z), and a first loss function between the first encoding E(Y) and the second encoding E(G(Z)) is minimized for the encoder; and
training the generator using the first encoding E(Y) and the second encoding E(G(Z)) as part of a second loss function for the generator.
9. The method of claim 8 , wherein a same set of weights are used for first encoding E(Y) of the target image Y and the second encoding E(G(Z)) of the generated image G(Z) corresponding to the target image Y, and the same set of weights are updated after one or more iterations of training the GAN.
10. The method of claim 8 , wherein the encoder comprises a machine learning model.
11. The method of claim 8 , wherein the encoder comprises a convolutional neural network.
12. The method of claim 8 , wherein the discriminator, the generator, and the encoder are trained, in sequential order, in each iteration of training the GAN.
13. The method of claim 8 , wherein for each iteration of training the GAN, the generator and the encoder are updated simultaneously after the discriminator has been updated.
14. The method of claim 8 , wherein the encoder is adapted from at least one of a first few layers of a VGG model, or one or more upscaling layers following the first few layers of the VGG model.
15. A method for image generation based on a Generative Adversarial Network (GAN), the GAN comprising a generator, a discriminator, and an encoder, wherein outputs of the generator are mapped, by the encoder, to a latent space adaptable to manipulate at least one characteristics of images generated by the GAN, the method comprising:
generating, by the encoder, a first encoding E(Y) of a target image Y and a second encoding E(G(Z)) of a generated image G(Z) corresponding to the target image Y, wherein the generated image G(Z) corresponding to the target image Y is generated by the generator, and the first and second encodings E(Y) and E(G(Z)) map the target image Y and the generated image G(Z) to the latent space having a lower dimension than dimensionality of the target image Y and the generated image G(Z), wherein the encoder is trained to minimize the differences between the first and second encodings E(Y) and E(G(Z)), and the generator is trained by using the first and second encodings E(Y) and E(G(Z)) as part of a loss function of the generator.
16. The method of claim 15 , wherein the discriminator is trained to distinguish which of the target image Y and the generated image G(Z) corresponding to the target image Y is real data.
17. The method of claim 15 , wherein a same set of weights are used for first encoding E(Y) of the target image Y and the second encoding E(G(Z)) of the generated image G(Z) corresponding to the target image Y, and the same set of weights are updated after one or more iterations of training the GAN.
18. The method of claim 15 , wherein the encoder comprises a machine learning model.
19. The method of claim 15 , wherein the encoder comprises a convolutional neural network.
20. The method of claim 15 , wherein at least one of:
the discriminator, the generator, and the encoder are trained, in sequential order, in each iteration of training the GAN; or
for each iteration of training the GAN, the generator and the encoder are updated simultaneously after the discriminator has been updated.