when the model is converted in the sdk, if both numParamBits and numFeatureBits in the import config are set to 8 bits, the model's inference results will differ significantly from those of the pc. howwever, if they are set to 16 bits, the model's inference results will be basically conssistent with those of pc.
for this purpose, we located the network layers that led to a significant drop in accuracy through a mixed-precision approach. these network layers are named in the.layer_info.txt file generated by th import tool: 37,38,39,40,41,195,196,197,198,199,204,266,267,268,269,270. these layers were transformed using 16-bit data precision to obtain a mixed-precision model. the inference results of the mixed_precision model have a significant improvement in precision compared to the 8-bit model. however, the inference speed of the mixed-precision model is much slower than that of the 8-bit model. our requirement is for fast inference speed and high precision
