An On-device Deep Neural Network for Face Detection - Apple

We implement several strategies to minimize memory footprint and GPU usage. To reduce memory footprint, we allocate the intermediate layers of our neural networks by analyzing the compute graph. This allows us to alias multiple layers to the same buffer. While being fully deterministic, this technique reduces memory footprint without impacting the performance or allocations fragmentation, and can be used on either the CPU or GPU. For Vision, the detector runs 5 networks (one for each image pyramid scale as shown in Figure 2). These 5 networks share the same weights and parameters, but have different shapes for their input, output, and intermediate layers. To reduce footprint even further, we run the liveness-based memory optimization algorithm on the joint graph composed by those 5 networks, significantly reducing the footprint.