Deep Quantization Generative Networks

Abstract Equipped with powerful convolutional neural networks (CNNs), generative models have achieved tremendous success in various vision applications. However, deep generative networks suffer from high computational and memory costs in both model training and deployment. While many efforts have been devoted to accelerate discriminative models by quantization, effectively reducing the costs for deep generative models is more challenging and remains unexplored. In this work, we investigate applying quantization technology to deep generative models. We find that keeping as much information as possible for quantized activations is key to obtain high-quality generative models. With this in mind, we propose Deep Quantization Generative Networks (DQGNs) to effectively accelerate and compress deep generative networks. By expanding the dimensions of the quantization basis space, DQGNs can achieve lower quantization error and are highly adaptive to complex data distributions. Various experiments on two powerful frameworks (i.elet@tokeneonedot, variational auto-encoders, and generative adversarial networks) and two practical applications (i.elet@tokeneonedot, style transfer, and super-resolution) demonstrate our findings and the effectiveness of our proposed approach.