Ram Cherukuri, MathWorks
With GPU Coder, you can deploy a deep neural network in MATLAB® to NVIDIA® Jetson™ board. You can either create a deep neural network and train it from scratch, or start with a pretrained network and retrain it through transfer learning. To learn more about this process, view the available resources on training a deep learning network in MATLAB.
CUDA® code can be generated from the neural network with GPU Coder™, along with the pre-processing and post-processing code that constitutes your MATLAB algorithm for an embedded vision application, for example. The generated CUDA code contains the binary weight and bias files for the various layers in the network.
You can then deploy your application, along with the deep learning network for inference, onto an embedded platform, such as NVIDIA Jetson TX1 board, by exporting the generated code to the target and building it on the target. Alternatively, you can also cross compile for the target on the host desktop.