经过训练得到最佳量化区间——Quantization Interval Learning

This article introduces a method for learning optimal quantizers through task loss optimization (QIL: Quantization Interval Learning). The approach effectively determines the best quantization intervals or quantizer parameters for network weights and activations. For a given training task, such as ImageNet image classification, the method first designs a trainable, parameterizable quantizer. During the training process (aiming to minimize task loss), both the floating-point weights of the neural network and the trainable parameters of the quantizer are updated. This results in a quantization s