Playground to try a basic machine learning model on a raspberry pi pico microcontroller with 2x Cortex M0+ cores @ 133Mhz, 264kB ram.

we will try out a simple pretrained fully connected model that maps an input float to a it's sine value, ref: https://www.youtube.com/watch?v=BzzqYNYOcWc

The model was loaded into netron, and the weights were extracted as .bin files then converted to hex-text form with a hex viewer (ref. src/parser.c for model structuring)

We'll also try implementing some rudimentary optimizations to optimize the model running on the microcontroller such as:

• brute force (overclocking the cpu core)
• multicore processing
• optimization for model sparsity

• You will need a Raspberry Pi pico and a work environment preferably setup, you can follow the documentation at: https://www.raspberrypi.com/documentation/microcontrollers/c_sdk.html

• once you've downloaded the pico's SDK you can clone this repository:

$ git clone https://github.com/ZephyrLabs/pico-ml

• setup CMake

$ cd pico-ml/src
$ mkdir build 
$ cd build
$ cmake -DPICO_SDK_PATH=<path-to-pico-sdk> ..

• build the program

$ make -j

• hold down the BOOTSEL button on your rpi pico and plugin the USB cable to your PC to put it into DFU mode.
• copy pico-delegate.uf2 from the build folder to the rpi pico's folder

• The device presents a simple serial console with a baud rate of 115200, connect to it with the serial console app of your choice. eg. on linux with picocom:

$ sudo picocom -b 115200 /dev/ttyACM0

Basic serial command guide:

• m: model parameters

  • l: load a byte of model data
  • r: reset model buffer
  • p: parse the model, verify structure

• t: tensor parameters:

  • l: load a byte of data into the tensor
  • d: dump the tensor buffer in the serial port, post inference
  • r: reset tensor buffer

• i: inferencing

  • i: run inference, if parsing was okay

eg: sending the following over serial

• load a byte into the model buffer: ml<some data byte>
• run inference: ii
• dump the tensor: td

• Model Params:
  • input_tensor_size: 4 bytes
  • output_tensor_size: 4 bytes
  • layer_count: 11
  • op_count: 66817 OPs
  • out: 0.845167

Total optimization: 12.08x

Note: these tests 1 - 4 involved running the model from flash memory which involves latency when fetching randomly from the XIP flash, executing from memory has been much faster.