Optimization technique for forming DNN capable of performing real-time inference in mobile environment
A system includes at least one processor having a learning part for learning a deep neural network (DNN)-based style transfer model by using an image of a specific style to be learned, the style transfer model being a DNN model having an architecture in which the number of deep layers is reduced through transfer learning using a previously learned result.
1. A system for optimizing a deep neural network (DNN) model, comprising:
at least one processor configured to execute computer-readable commands, the processor including,
a learning part for learning a deep neural network (DNN)-based style transfer model by using an image of a specific style to be learned,
wherein the style transfer model is a DNN model having an architecture in which the number of deep layers is reduced through transfer learning using a previously learned result and includes a plurality of layers of a previously trained DNN model that has learned an image of a style similar to the specific style,
wherein the previously trained DNN model is a selected DNN model from a list including a plurality of previously trained DNN models based on style similarities between the style transfer model to be trained and each of the plurality of previously trained DNN models and
wherein the plurality of previously trained DNN models are trained based on a style of painting of an image by using a gram matrix.
2. The system of claim 1 , wherein the style transfer model is a DNN model having an architecture in which a feature size of at least one layer is reduced.
3. The system of claim 1 , wherein the style transfer model is a DNN model having an architecture in which an instance normalization operator is added to a plurality of residual layers.
4. The system of claim 3 , wherein each of the residual layers includes a first convolution operator, a first instance normalization operator, an activation function, a second convolution operator, and a second instance normalization operator, in order.
5. The system of claim 3 , wherein a last residual layer of the plurality of residual layers comprises an architecture for scaling down a resultant value of an operation of a preceding layer.
6. The system of claim 1 , wherein the at least one processor including the learning part is provided in a mobile electronic device.
7. The system of claim 6 , wherein the at least one processor includes a half-precision type processor, and the plurality of previously trained DNN models are configured to operate on a single-precision type processor or a double-precision type processor.
8. A method, executed by a computer, for optimizing a deep neural network (DNN) model, comprising:
learning a deep neural network (DNN)-based style transfer model by using an image of a specific style to be learned; and
providing a resulting image by applying the specific style to an input image through the style transfer model,
wherein the style transfer model is a DNN model having an architecture in which the number of deep layers is reduced through transfer learning using a previously learned result and includes a plurality of layers of a previously trained DNN model that has learned an image of a style similar to the specific style,
wherein the previously trained DNN model is a selected DNN model from a list including a plurality of previously trained DNN models based on style similarities between the style transfer model to be trained and each of the plurality of previously trained DNN models, and
wherein the plurality of previously trained DNN models are trained based on a style of painting of an image by using a gram matrix.
9. The method of claim 8 , wherein the style transfer model is a DNN model having an architecture in which a feature size of at least one layer is reduced.
10. The method of claim 8 , wherein the style transfer model is a DNN model having an architecture in which an instance normalization operator is added to a plurality of residual layers.
11. The method of claim 10 , wherein each of the residual layers includes a first convolution operator, a first instance normalization operator, an activation function, a second convolution operator, and a second instance normalization operator, in order.
12. The method of claim 10 , wherein a last residual layer of the plurality of residual layers comprises an architecture for scaling down a resultant value of an operation of a preceding layer.
13. A computer-readable recording medium storing a program for instructing a computer to execute the method for optimizing a deep neural network model as defined in claim 8 .