implementation issue

[hidebusher]

Hi, first, thanks for your implementation!
In the original paper, it is mentioned that weights and activations are calculated by integers during inference.However, in your training code,weights and activations are still floating-point numbers after pseudo quantization.Should they be multiplied by integers during training?Otherwise, how to ensure the accuracy of integer multiplication and then quantifying them back?And how can the inference time be reduced when using floating-point numbers for calculations?