Performance of Post-Training Two-Bits Uniform and Layer-Wise Uniform Quantization for MNIST Dataset from the Perspective of Support Region Choice

Table 2

SQNR and model accuracy for application of different two-bits LWUQ designs.

Two-bits LWUQ, accuracy (FP32) = 98.1%
(xmax^L1, xmax^L2, xmax^L3) = (4.5150, 4.8371, 3.6784); (xmin^L1, xmin^L2, xmin^L3) = (−7.0638, −5.4354, −6.1979)	Case 1 ℜg	Case 2ℜg
	[−xmax^L1, xmax^L1]; [−xmax^L2, xmax^L2]; [−xmax^L3, xmax^L3]	[xmin^L1, −xmin^L1]; [xmin^L2, −xmin^L2]; [xmin^L3, −xmin^L3]
SQNRex^UQ^L1 (dB)	3.1340	−1.7588
SQNRex^UQ^L2 (dB)	3.4507	2.2826
SQNRex^UQ^L3 (dB)	8.3642	4.6137
SQNRex^LWUQ (dB)	3.3145	−0.374
Accuracy (%)	97.26	93.55