Three-color face dynamic recognition (virtual AR) 1, first make 2, then convert 3, then manually control, installation method, the compressed package of each app should be right-clicked to control decompression, and then the portrait app prompts to install twice. The conversion app can be installed normally, and the control app does not need to be installed. Finally, watch the tutorial video of MP4 first, and then try it slowly.
Network model quantization, ip unit with int8 operation is built in equipment with NPU, and post-training quantization toolbox is generally provided for users to convert models. Similarly, Ruixinwei also provides the corresponding tool chain, which has been opened on github, rknn-toolkit.
Generally, we only need to pay attention to the following points when quantifying the algorithm network: 1, and the provided calibrationdataset should be distributed with the training data as much as possible, or simply extract some data from the training data for calibration. Generally, the legend of 1k is provided for post-training quantization, which can obtain better numerical accuracy. Some ramblings about quantification have also been discussed with you in previous articles. You can refer to MNN's model to quantify ① the mathematical model.
2. In NPU project (rk, allwin, k2 10, etc. ), try to kill DW-CONV when designing the network. Op is not well adapted in NPU, and the quantization accuracy of op under int8 is often overturned (high-level occasionally overturned, low-level high probability overturned). I usually use resnet-like modules for miniaturization, and the speed and performance are generally better than those based on mobilenet structure.
Tri-color world is an auxiliary tool for face recognition, which is face recognition through human face.