It would be nice if you could cope with quantized tensors. i.e. Allow fixed point 64 32 16 and 8 bit ints. Add a normalization field to the tensor for these types.
Add FIXP64, FIXP32, FIXP16 and FIXP8 types and manditory normalization field for these types.
operator
sousoux
👍1
All 3 comments
It would also be nice to allow for flexible enough linear quantization schemes, allowing arbitrary scale (I beleive what is referred to above as "normalization") and offset to allow for non-symmetric dynamic range.
sergeigofman
on 17 Dec 2017
👍4
houseroad
on 15 Mar 2018
https://github.com/onnx/onnx/wiki/Quantization-Support-In-ONNX shows the principals that are followed when adding quantization support in onnx. The first small set of quantized operators were merged to support linear 8-bit quantization. More may be added as needed.
linkerzhang
on 19 Jul 2019
👍1
Was this page helpful?
0 / 5 - 0 ratings
Related issues
fdwr
路
3Comments
ezyang
路
4Comments
chiehpower
路
4Comments
yetingqiaqia
路
3Comments
cdeterman
路
5Comments
Most helpful comment
It would also be nice to allow for flexible enough linear quantization schemes, allowing arbitrary scale (I beleive what is referred to above as "normalization") and offset to allow for non-symmetric dynamic range.