Post-training Quantization with Multiple Points: Mixed Precision without Mixed Precision

We consider the post-training quantization problem, which discretizes weights of pre-trained deep neural networks without re-training model. propose multipoint quantization, a method that approximates full-precision weight vector using linear combination multiple vectors low-bit numbers; this is in contrast to typical methods approximate each single low precision number. Computationally, we construct with an efficient greedy selection procedure, and adaptively decides number points on quantized based error its output. This allows us achieve higher levels for important greatly influence outputs, yielding ``effect mixed precision'' but physical implementations (which requires specialized hardware accelerators). Empirically, our can be implemented by common operands, bringing almost no memory computation overhead. show outperforms range state-of-the-art ImageNet classification it generalized more challenging tasks like PASCAL VOC object detection.