Automated compilation of deep learning neural networks for image processing

In recent years, machine learning has very much been a prominent talking point, and is considered by many to be the next big thing, comparable to, if not greater than the dot-com boom of the late 90s. Whilst machine learning and AI are general terms, an area that has seen significant growth both in research and industrial or commercial applications are neural networks, especially those known as deep neural networks, where the increases in processing power available through higher core count CPUs and mutliple high power GPUs have meant that networks can now be far bigger than when they were originally put forward.

This work looks at a subsection of this field, focusing on network inference instead of training, and being able to utilise trained networks on multiple platforms as easily as possible. Due to the high growth of the area, there are many specialist tools available to allow users to use neural networks on a wide range of hardware, however in order to test and develop on multiple platforms such as x86 and embedded or FPGA systems, you are required to use multiple tools, each separate from one another. This work aims to provide a single solution allowing a user to test trained networks on multiple platforms, knowing the code base originates from the same core and lowering some of the difficulties that embedded or FPGA platforms can throw up when it comes to developing for them.

The system created allows a user to target multiple different platforms, automatically generating code to be used on each, reducing the number of steps required before the network can be up and running on the chosen device. TinyYOLO was the main network used for testing, on multiple x86 platforms, and a Digilent Zedboard for freeRTOS, PetaLinux and FPGA testing.