Jinbae Park

@zhutmost Can you share your experimental results on ImageNet? I wonder the bit-width and its accuracy.

@zhutmost Thank you! I'm also trying to improve the LSQ (using Tensorflow), since I got 70.7% top-1 accuracy for the W4A4 case. If you figure out some detailed hyper-parameter effects,...

@yuslepukhin Yes, sure. This is the entire log. ``` 2022-08-12 12:52:57.4010505 [I:onnxruntime:, inference_session.cc:262 onnxruntime::InferenceSession::ConstructorCommon::::operator ()] Flush-to-zero and denormal-as-zero are off 2022-08-12 12:52:57.4154053 [I:onnxruntime:, inference_session.cc:270 onnxruntime::InferenceSession::ConstructorCommon] Creating and using per session...

According to the Eq. (5) in the LSQ paper, the derivative of v is limited by the range. Thus, I think the backward function can be reimplemented as follow: ```...

Thank you! Actually, I used tensorflow & tensorpack frameworks. I also wonder about the results of pytorch.