Generating images using neural networks
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating images using neural networks. One of the methods includes generating the output image pixel by pixel from a sequence of pixels taken from the output image, comprising, for each pixel in the output image, generating a respective score distribution over a discrete set of possible color values for each of the plurality of color channels.
1. A computer-implemented method of generating an output image comprising a plurality of pixels arranged in a two-dimensional map, each pixel having a respective value for each of one or more channels, and wherein the method comprises:
receiving a neural network input by a neural network system;
processing the neural network input using one or more initial neural network layers of the neural network system to generate a respective portion of an alternative representation of the neural network input corresponding to each of the one or more channels for each pixel in the plurality of pixels included in the output image; and
for each of the one or more channels for each pixel in the plurality of pixels included in the output image:
processing the respective portion of the alternative representation using one or more output neural network layers to generate a respective score distribution over a discrete set of possible values for the channel, wherein the respective score distribution comprises a score for each of the possible values in the discrete set; and
selecting, using the respective score distribution and from the discrete set of possible values for the channel, the respective value for the channel.
2. The computer-implemented method of claim 1 , wherein generating the output image comprising the plurality of pixels arranged in the two-dimensional map comprises generating the output image pixel by pixel from a sequence of pixels taken from the output image.
3. The computer-implemented method of claim 2 , wherein processing the neural network input using the one or more initial neural network layers comprises applying a mask to outputs of neural network layers in the one or more initial neural network layers such that the respective portion of the alternative representation corresponding to the channel is only generated based on (i) values for pixels before the pixel in the sequence and (ii) values for the pixel for any channels before the channel in an order of channels.
4. The computer-implemented method of claim 3 , wherein the one or more initial neural network layers comprise one or more convolutional neural network layers.
5. The computer-implemented method of claim 4 , wherein the mask is implemented as a convolutional layer kernel having zero values corresponding to pixels after the pixel in the sequence.
6. The computer-implemented method of claim 4 , wherein the one or more initial neural network layers comprise a horizontal stack of convolutional neural network layers and a vertical stack of convolutional neural network layers, wherein, for a given pixel in a given row, the horizontal stack conditions on the values already generated for the given row so far and the vertical stack conditions on all rows above the given row.
7. The computer-implemented method of claim 4 , wherein the alternative representation is a feature map generated by a convolutional neural network layer that includes features for each of the one or more channels for each pixel in the plurality of pixels included in the output image.
8. The computer-implemented method of claim 1 , wherein the neural network input comprises an image.
9. The computer-implemented method of claim 1 , wherein the one or more output neural network layers comprise one or more softmax layers each configured to generate a respective score for each integer value in a set of integer values representing a likelihood that the integer value should be the value of a given channel.
10. A system comprising one or more computers and one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations for generating an output image comprising a plurality of pixels arranged in a two-dimensional map, each pixel having a respective value for each of one or more channels, and wherein the operations comprise:
receiving a neural network input by a neural network system;
processing the neural network input using one or more initial neural network layers of the neural network system to generate a respective portion of an alternative representation of the neural network input corresponding to each of the one or more channels for each pixel in the plurality of pixels included in the output image; and
for each of the one or more channels for each pixel in the plurality of pixels included in the output image:
processing the respective portion of the alternative representation using one or more output neural network layers to generate a respective score distribution over a discrete set of possible values for the channel, wherein the respective score distribution comprises a score for each of the possible values in the discrete set; and
selecting, using the respective score distribution and from the discrete set of possible values for the channel, the respective value for the channel.
11. The system of claim 10 , wherein generating the output image comprising the plurality of pixels arranged in the two-dimensional map comprises generating the output image pixel by pixel from a sequence of pixels taken from the output image.
12. The system of claim 11 , wherein processing the neural network input using the one or more initial neural network layers comprises applying a mask to outputs of neural network layers in the one or more initial neural network layers such that the respective portion of the alternative representation corresponding to the channel is only generated based on (i) values for pixels before the pixel in the sequence and (ii) values for the pixel for any channels before the channel in an order of channels.
13. The system of claim 12 , wherein the one or more initial neural network layers comprise one or more convolutional neural network layers.
14. The system of claim 13 , wherein the mask is implemented as a convolutional layer kernel having zero values corresponding to pixels after the pixel in the sequence.
15. The system of claim 13 , wherein the one or more initial neural network layers comprise a horizontal stack of convolutional neural network layers and a vertical stack of convolutional neural network layers, wherein, for a given pixel in a given row, the horizontal stack conditions on the values already generated for the given row so far and the vertical stack conditions on all rows above the given row.
16. The system of claim 13 , wherein the alternative representation is a feature map generated by a convolutional neural network layer that includes features for each of the one or more channels for each pixel in the plurality of pixels included in the output image.
17. The system of claim 10 , wherein the neural network input comprises an image.
18. The system of claim 10 , wherein the one or more output neural network layers comprise one or more softmax layers each configured to generate a respective score for each integer value in a set of integer values representing a likelihood that the integer value should be the value of a given channel.
19. One or more non-transitory computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for generating an output image comprising a plurality of pixels arranged in a two-dimensional map, each pixel having a respective value for each of one or more channels, and wherein the operations comprise:
receiving a neural network input by a neural network system;
processing the neural network input using one or more initial neural network layers of the neural network system to generate a respective portion of an alternative representation of the neural network input corresponding to each of the one or more channels for each pixel in the plurality of pixels included in the output image; and
for each of the one or more channels for each pixel in the plurality of pixels included in the output image:
processing the respective portion of the alternative representation using one or more output neural network layers to generate a respective score distribution over a discrete set of possible values for the channel, wherein the respective score distribution comprises a score for each of the possible values in the discrete set; and
selecting, using the respective score distribution and from the discrete set of possible values for the channel, the respective value for the channel.
20. The non-transitory computer-readable storage media of claim 19 , wherein generating the output image comprising the plurality of pixels arranged in the two-dimensional map comprises generating the output image pixel by pixel from a sequence of pixels taken from the output image.