Add bias correction

[Jiang-Stan]

• Backend: virtual
• W: Symmetric
  • ResNet18: MSE Observer
  • MobilenetV2: MinMax Observer
  • DeiT: MinMax Observer
• A: layerwise 8bit, MinMax Observer, Symmetric
• Eval num: 5000

Model	configs	weight granularity	top1	top5	reported top1(from Adaround paper)
ResNet18	float	-	69.86	89.4
ResNet18	4w8f	layerwise	20.16	40.12	23.99
ResNet18	4w8f w/ BC	layerwise	38.04	62.44	38.87
ResNet18	4w8f	channelwise	53.86	76.02	-
ResNet18	4w8f w/ BC	channelwise	59.60	82.96	-
DeiT	float	-	72.76	91.12	-
DeiT	8w8f	channelwise	71.52	90.76	-
DeiT	8w8f w/ BC	channelwise	71.58	90.42	-
DeiT	4w8f	channelwise	64.18	86.42	-
DeiT	4w8f w/ BC	channelwise	64.68	86.82	-

Model	configs	weight qscheme	feature qscheme	top1	top5
MobileNetV2	float	-	-	72.60	90.10
MobileNetV2	8w8f	per-tensor-symmetric	per-tensor-symmetric	67.56	87.44
MobileNetV2	8w8f w/ BC	per-tensor-symmetric	per-tensor-symmetric	70.10	88.98
MobileNetV2	8w8f	per-tensor-symmetric	per-tensor-affine	68.86	88.46
MobileNetV2	8w8f w/ BC	per-tensor-symmetric	per-tensor-affine	71.32	89.96
MobileNetV2	8w8f	per-channel-symmetric	per-tensor-symmetric	70.50	89.50
MobileNetV2	8w8f w/ BC	per-channel-symmetric	per-tensor-symmetric	70.58	89.44
MobileNetV2	8w8f	per-channel-symmetric	per-tensor-affine	71.44	89.64
MobileNetV2	8w8f w/ BC	per-channel-symmetric	per-tensor-affine	72.1	89.82

[hych2020]

加个mobilenetv2的bias correction结果(8w8f)