A simple logo classifier developed using Maixduino framework and PlatfomIO, to run on K210 MCU on Sipeed's Maix dev board. I trained my own ML model, using transfer learning from MobileNet v1.

Click the thumbnail

• PlatformIO
• platform-kendryte210. Should be installed automatically
• Kendryte nncase for NeuralNet optimization, download from here. Unzip anywhere.
• If you're like me, I'll use VSCode and install PlatformIO extension. Maixduino is available for Arduino IDE, but real programmer knows what they should use.

• Install Tensorflow, Keras, and other stuffs. RTFM.
• As the trained model leverages MobileNet, apparently we need to adjust it to be compatible with K210. Replace mobilenet.py file on site-packages/keras_applications (don't forget to backup) with the one in this repo. site-packages folder may exist on several places depends on your environment. If you use virtualenv, it should be under you_virtualenv_dir/lib/python3.x
• Take a look at training/mbnet_keras.py file. Adjust the constants, and run it.
• Convert the generated h5 model file by running training/convert.sh script with the h5 model file as parameter. Eg. ./convert.sh logoclassifier.h5
• Copy the generated kmodel file to src
• Adjust the labels on src/names.cpp file

(More complete steps will be coming soon)

• Some code and steps are inspired by this useful tutorial. Thanks for your support @AIWintermuteAI
• MobileNet class is adapted from MBNet_1000 class from Maixduino example