OFMR Algorithm

The Output Feature Map Reconstruction (OFMR) algorithm determines the optimal quantization factors by searching in an activation distribution. This algorithm applies to PTQ. The following figure shows the quantization principles.

1. Infer the original model to obtain the source output of the model to be quantized.

2. Transverse all candidate quantization factors, infer the model, and obtain the quantized output.
3. Compare the quantized output with the source output to select the quantization factor with the smallest error.