CodePudding user response:
Using SNPE deployment model of general working process is divided into the following figure described in two stages: the first stage by the Caffe/Caffe2/TensorFlow complete model of training; The second phase will be deployed to qualcomm device model; Before deployment model with SNPE tools, transform the model for SNPE unique DLC (Deep Learning Container) file format, can be after conversion Xiao dragon CPU/GPU performs reasoning on the DSP core, specific steps are as follows:1, transform the model into SNPE support of DLC format;
2, quantification is optional, if deployed to the Runtime or CPU - FXP Hexagon DSP CPU fixed-point, need to do the model offline eight fixed-point quantitative (DSP can also be removed from the offline quantization steps), offline quantitative need to specify the data;
3, for the model is used to test the input data: use./models/alexnet/scripts/script, convert JPG image file to raw format, generating raw_list. TXT file (note: SNPE input data format is BGR, and each pixel value BGR is continuous, namely NHWC format,