WQ的max_val为什么是 (1 << self.bits) - 2)？

[xingyueye]

你好，我注意到IAO的quantizer中，weights的max_val被设置为 (1 << self.bits) - 2 ，而activation是(1 << self.bits) - 1)。不知道为什么这样设置，两者不应该都是255（8bit）吗?

class UnsignedQuantizer(Quantizer):
    def __init__(self, *args, **kwargs):
        super(UnsignedQuantizer, self).__init__(*args, **kwargs)
        if self.activation_weight_flag == 0:  # weight
            self.register_buffer(
                "quant_min_val", torch.tensor((0), dtype=torch.float32)
            )
            self.register_buffer(
                "quant_max_val",
                torch.tensor(((1 << self.bits) - 2), dtype=torch.float32),
            )
        elif self.activation_weight_flag == 1:  # activation
            self.register_buffer(
                "quant_min_val", torch.tensor((0), dtype=torch.float32)
            )
            self.register_buffer(
                "quant_max_val",
                torch.tensor(((1 << self.bits) - 1), dtype=torch.float32),
            )
        else:
            print("activation_weight_flag error")