How QuantHardTanh works

[wanglu4042]

Hi, I have a question for QuantHardTanh.
I set it like the picture downbelow.

but after I saved the module as onnx, I found that QuantHardTanh was changed like this.

And in FINN, it was like this:

And all parameters were set to float32, especially the part called MultiThreshold, it had the parameters like this:

Could you please tell me, how are these numbers calculated when they are INT8 and also in float32?
Thanks!!
Looking forward to your reply!

[Giuseppe5]

Hi,

Generally speaking, it is always useful to have a minimal script to reproduce your observations, including versions of the packages involved.

Many thanks!