search

huawei-noah / AdderNet

Code for paper " AdderNet: Do We Really Need Multiplications in Deep Learning?"
BSD 3-Clause "New" or "Revised" License
954 stars 187 forks source link

AdderNet for SR #42

Closed leiwen83 closed 1 year ago

leiwen83 commented 3 years ago

Hi,

I see recent you have applying addernet for EDSR quantization, which is a great work. But for some other SR, like EDVR, using deformable conv for the frame alignment.

I wonder whether addernet also be efficient, or even can be applied to this kind of network? Since the DCNv2 using in EDVR is a irregular conv, how to adjust quantization int8 loss introduced by adderNet? Do you have some thought over this kind of op?

Thx, Lei

HantingChen commented 3 years ago

We have not try to replace the deformable conv by adder operation. AdderNet can also be efficient for this kind of network if successfully applied.

For the int8 quantization, we use the post-quantization (no retrain, no loss) and you may refer to https://arxiv.org/pdf/2101.10015 for more information.