Abstract
Training of convolutional neural networks (CNNs) on embedded platforms to support on-device learning has become essential for the future deployment of CNNs on autonomous systems. In this work, we present an automated CNN training pipeline compilation tool for Xilinx FPGAs. We automatically generate multiple hardware designs from high-level CNN descriptions using a multi-objective optimization algorithm that explores the design space by exploiting CNN parallelism. These designs that trade-off resources for throughput allow users to tailor implementations to their hardware and applications. The training pipeline is generated based on the backpropagation (BP) equations of convolution which highlight an overlap in computation. We translate the overlap into hardware by reusing most of the forward pass (FP) pipeline reducing the resources overhead. The implementation uses a streaming interface that lends itself well to data streams and live feeds instead of static data reads from memory. Meaning, we do not use the standard array of processing elements (PEs) approach, which is efficient for offline inference, instead we translate the architecture into a pipeline where data is streamed through allowing for new samples to be read as they become available. We validate the results using the Zynq-7100 on three datasets and varying size architectures against CPU and GPU implementations. GPUs consistently outperform FPGAs in training times in batch processing scenarios, but in data stream scenarios, FPGA designs achieve a significant speedup compared to GPU and CPU when enough resources are dedicated to the learning task. A 2.8x, 5.8x, and 3x speed up over GPU was achieved on three architectures trained on MNIST, SVHN, and CIFAR-10 respectively.