Quantitative perception training in tensorflow2.x and x86 end evaluation of tflite

tensorflow2.x Quantitative perception training and tflite Of x86 End evaluation _BIT_Legend The blog of -CSDN Blog _tensorflow Quantitative perception training tensorflow 2.x After the model training, it is generally saved as .h5 or save_model The model format of ( Can only be fp32 Format ), If the model wants to be deployed on the mobile end , Generally, you need to convert the model into .tflite Format , Here it is divided into dynamic format 、fp32 Format 、fp16 Format and int8 Format tflite【1】. generally speaking , If you will fp32 Of h5 Convert to int8 Format tflite There will be a decrease in accuracy , In order to minimize the accuracy decline in the conversion process , The optional way is to train the model with quantitative perception before transformation . The specific method is to maintain the weight fp32 Under the circumstances , according to int8 Quantify the need for training , Then it will be officially carried out inhttps://blog.csdn.net/BIT_Legend/article/details/122266428