Quantization of Neural Networks for Efficient Inference

The expressiveness of modern Generative Adversarial Networks is connected to massive amounts of computations performed during the inference stage, complicating their deployment on edge de­ vices. On the other hand, quantization is a popular and effective neural network compression tech­ nique that allows to perform hardware­friendly inference. However, while recent works indicate the possibility to quantize discriminative models to a low number of bits even with uniform schemes, the performance of modern quantization techniques in application to GANs remains unclear. This paper tackles the considered problem by conducting an extensive experimental study of state­of­ art quantization techniques’ effectiveness to uniform quantization of Generative Adversarial Net­ works. The study concerns post­training quantization and quantization­aware training techniques applied to generators of three diverse GAN architectures, namely StyleGAN, Self­Attention GAN, and CycleGAN. The results suggest that these models can be successfully quantized for 4/8­bit inference with negligible quality degradation.