Venturi is a lightweight, configurable neural net library developed using Rust. The library has been developed to provide a starting point for adding basic AI capabilities to low-cost embedded systems.
It was originally inspired by the neural net code presented in the book, "Make Your Own Neural Network" by Tariq Rashid.
The MNIST (Modified National Institute of Technology database) database is used both in the book and to exercise the original Rust implementation of the neural net. Small samples of the MNIST database are included that can be used both for training and query experiments.
Venturi is currently using capabilities that would typically require a hosted
environment. Future development will support a #![no_std] option for bare
metal environments. See Rust Embedded
for additional information.
The project includes a simple baseline benchmark of the training method using a fixed dataset. The benchmark feature needs to be enabled to run the tests and you must use the nightly toolchain:
% rustup override set nightly
% cargo bench --tests --features benchmark
The benchmark baselines are taken from the average ns/iter ovber 5 benchmark runs on each platform.
| Platform | OS | CPU | Memory | Benchmark |
|---|---|---|---|---|
| MacBook Pro 15" 2019 | 12.2.1 | 2.6GHz Intel Core i7 | 16GB 2400 MHz DDR4 | 26_772 ns/iter +/- 1661 |
| Raspberry Pi 4 | Raspbian/GNU/Linux 11 | Cortex-A72 (ARM v8) 64-bit 1.5GHz | 4GB LPDDR4-3200 SDRAM | 72_602 ns/iter +/- 1190 |
The project includes a CLI entry point that is able to exercise the library capabilities.
You can see the command line help by running venturi either through the cargo run command
cargo run --release -- --help
or by running venturi directly
venturi --help
The mode argument must be one of:
- train
- query
The input-file argument is used in query mode and the training-data argument is used in training mode.
The output-file argument may be used to create a binary file of the trained network for future queries.
USAGE:
venturi [FLAGS] [OPTIONS] --mode <mode>
FLAGS:
-h, --help Prints help information
-s, --show-training
-V, --version Prints version information
OPTIONS:
-H, --hidden-node-count <hidden-node-count>
-i, --input-file <input-file>
-I, --input-node-count <input-node-count>
-m, --mode <mode>
-n, --network-file <network-file>
-o, --output <output>
-O, --output-node-count <output-node-count>
-t, --training-data <training-data>