A plethora of applications are using machine learning, the operations of which are becoming more complex and require additional computing power. At the same time, typical commodity system setups (including desktops, servers, and embedded devices) are now offering different processing devices, the most often of which are multi-core CPUs, integrated GPUs, and discrete GPUs. In this paper, we follow a data-driven approach, where we first show the performance of different processing devices when executing a diversified set of inference engines; some processing devices perform better for different performance metrics (e.g., throughput, latency, and power consumption), while at the same time, these metrics may also deviate significantly among different applications. Based on these findings, we propose an adaptive scheduling approach, tailored for machine learning inference operations, that enables the use of the most efficient processing device available. Our scheduler is device-agnostic and can respond quickly to dynamic fluctuations that occur at real-time, such as data bursts, application overloads and system changes. The experimental results show that it is able to match the peak throughput, by predicting correctly the optimal processing device with an accuracy of 92.5%, with energy savings up to 10%.