A Neural Network library that does not make use of allocations or the standard library at all. It does all its work on the stack.
This has some advantages:
Embedded devices without any operating system are now able to run at least simple neural networks.
Since the whole network layout needs to be known at compile time the dimensions of inputs and outputs are checked.
No need for OpenCL or CUDA, it just runs on your CPU. Or basically any other CPU for that matter.
The whole network being known to the compiler might enable some optimizations. That said the library is currently not very well optimized.
Also it was a fun challenge and actually worked out :)
Check the examples directory for some simple networks to get started.
- a convolutional layer would be nice
- think about the library design, specifically Layer might be too coarse of a trait, sub-layers may be useful.
- unify SoftMax and other layers
- better optimization
- ! move to less horrible generics
- figure out how to use less type parameters
This crate will be 1.0 if it has the tools to detect handwriting and is kinda easy to use.
Please symlink the hooks to your local .git/hooks/ directory to run some automatic checks before committing.
ln -s ../../hooks/pre-commit .git/hooks/
Please install rustfmt and cargo-sync-readme so these checks can be run.
rustup component add rustfmt
cargo install cargo-sync-readme
Please execute cargo-sync-readme when you change the top-level-documentation.
Please run cargo fmt whenever you change code. If possible configure your editor to do so for you.