CodePudding user response:
NCNN:1. NCNN support mixed precision, if not choose requantize option, the output of each layer will read the data to do quantitative to float
2. NCNN USES quantization algorithm, the feature map quantitative KLD algorithm is applied to make saturated processing (from TensorRT), the weights using channel quantization (for some of DW CONV network has a lot to improve)
3. Because NCNN mixing precision, actually avoid the shortcomings of some KLD algorithm, such as in the last few layer of the convolution is not suitable to do quantitative saturation
TensorFlow lite:
1. Not used in the quantitative saturation algorithm, is to collect each layer of min - Max do quantitative data, but the activation of the model value distribution demand higher (should be more compact, no obvious long tail)
2. Tensorflow by adding fake - quant node in the fine turning time constraints, the weight and the distribution of the activation values, reached the requirements of the first point,
CodePudding user response:
Can refer to this link: https://zhuanlan.zhihu.com/p/74326567