MNIST inference on the "3 cent" Microcontroller.

This project attempts to push BitNetMCU to the absolute limit: Is it possible to implement reasonably accurate inference of MNIST, the handwritten numbers dataset, on a "3 cent" Microcontroller with only 64 bytes of RAM and 1K of instruction memory?

Only weeks ago, my resounding answer would have been: "NO". But after the surprisingly good performance of BitNetMCU, I decided to give it a try.

This code achieves 90.1% test accuracy on test images scaled to 8x8 pixels while consuming 59 of 64 bytes of RAM and 972 of 1024 words of program memory. See blog for a few more details. The training code can be found in this branch of BitNetMCU

Project structure

├── include/                    # Contains the free-pdk header files
└── src/                
    ├── BitNetMCU_model.h       # MNIST model definition 
    ├── main.c                  # Main source file
    ├── Makefile
    ├── PDK_softuart.c          # Software UART implementation
    └── softuart.h              # Software UART header file

Building Instructions

!!! Use SDCC 4.1.0 to build. Later versions create broken code !!!

This is a hack, don't expect it to be useful for anything.

If you still want to build the project, you need to have the FreePDK toolchain installed. The default target is the PFS154, but you can change it to the PMS150C by editing the makefile. Currently, monitoring output is only supported with PFS154 as a target and the UART functions will not be included if PMS150C is selected.

1. Navigate to the src/ directory.
2. Run make flash to build and flash the project
3. make start to start the monitor