Recognizing "Hand-Drawn" Numbers on PyGamer via @DanielPBank

I'm working on making an example app for the PyGamer which allows the user to draw numbers and the app infers what number is drawn. I want to use a model that was trained on the MNIST dataset, load it in the app, and use it to run inference. I found a crate called tch, which is Rust bindings for the PyTorch C++ API. There is some example code in a GitHub Issue for saving / loading models. Now I just need to build it for an embedded system. Is it possible? I can add the #![no_std] pragma and it still builds. When I add the #![no_main] pragma, the build fails with an error:

error: linking with `cc` failed: exit code: 1
  = note: ld: entry point (_main) undefined. for architecture x86_64
          clang: error: linker command failed with exit code 1 (use -v to see invocation)

It may be possible to build this for an embedded system, but I need to inspect the source code thoroughly for things like usage of Vec or HashMap (or any kind of std::collections).

Jacob mentioned that he was working on a tensorflow-lite-sys, Rust bindings of the TensorFlow Lite for Microcontrollers. This would support embedded systems and may be an easier avenue to go.

Intro to Embedded via @craig_jbishop

Craig takes us through an Intro to Embedded Rust. Some highlights:

  • The various layers of the Rust embedded ecosystem: low-level cortex-m support, minimal cortex-m startup and runtime (cortex-m-rt), panic handlers, peripheral access crates (PACs), hardware access layers (HALs), peripheral and device drivers, board support crates, real time operating systems (RTOS), and finally application code.

  • The benefits of owning and using a debugger over using the bootloader mode.

  • Automating task running with a Makefile.toml and cargo-make

More Rust Embedded via @jacobrosenthal

Jacob expands upon Craig's talk with a few more points:

  • cortex-m-quickstart, a template for building apps for ARM Cortex-M microcontrollers

  • Using cargo-expand to expand the macro code of a Hello World Rust application (the app template resulting from running cargo new foo). There is still a lot of underlying code that depends on std libraries, none of which will work in an embedded application.

A Crate You Should Know via @PhilGebhardt

Phil talks about clap, a crate for doing command-line argument parsing. Jacob adds as an addendum the structopt crate, which is built on clap and uses derived macros.

How To Write Fast Rust Code via @ChristopherSebastian

Christopher shows off his extremely fast crate for evaluating algebraic expressions and explains some of the techniques he used to make it fast. He wrote a blog post about writing fast rust code and goes through the points, such as taking noise-free measurements, profiling with perf, compiling with RUSTFLAGS="--emit=asm", and reducing redundant computations.


Crates You Should Know

  • fasteval: Fast and safe evaluation of algebraic expressions
  • flamegraph: A simple cargo subcommand for generating flamegraphs, using inferno under the hood
  • tch: Rust wrappers for the PyTorch C++ api (libtorch)
  • structopt: Parse command line argument by defining a struct
  • clap: A simple to use, efficient, and full-featured Command Line Argument Parser