Automating ai model inference deployment to FPGA workflows

Machine learning is a powerful technique for many different data processing applications. Once a model is trained, it is typically deployed using a GPU, which is sufficient for many applications. However, it is advantageous in some cases to deploy an FPGA for greater throughput and lower latency. Creating an FPGA version of a given model is usually a difficult manual process, motivating the development of a tool to automate this process. Utilizing the popular machine learning framework PyTorch to generate models, I have created a tool which can quickly and easily translate a model into Vitis HLS, which allows for easy programming of FPGAs. The tool will handle the structure of the model as well as handling all of the parameters of the model. By doing this the model can be quickly and easily deployed to an FPGA accelerator.