Deep neural networks in Rust
Switch branches/tags
Nothing to show
Clone or download
Permalink
Failed to load latest commit information.
examples Having fully connected layer after LSTMs for char-rnn model was kinda… Jul 1, 2016
src Move 'extern crate num' declaration to lib.rs. Remove now-unused zero… Aug 20, 2017
.gitignore Added mnist example Feb 10, 2016
Cargo.toml Zero, One traits have been deprecated in Rust, use Num crate instead Aug 18, 2017
LICENSE Added MIT license Jan 20, 2016
README.md Updated readme Jun 17, 2016
download_mnist.sh

README.md

deeplearn-rs

Deep learning in Rust! This is my first shot at this. It's mostly just a proof of concept right now. The API will change.

Status

We have these models implemented (check out the examples folder):

  • MNIST handwritten digit recognition
  • char-rnn using LSTM

So far, we have the following layers implemented:

  • Matrix multiply (fully connected)
  • Add (for bias, for example)
  • LSTM
  • Softmax
  • MSE loss
  • Cross entropy loss

We have the following optimizers:

  • SGD
  • RMSProp

Road map

  • More layer types (in the order that I'll probably get to them)
    • Conv2d
    • Pooling
    • Dropout
  • Allow datatypes other than f32 and implement casting between arrays of primitive numeric types.
  • Provide utilities for working with data
    • images
    • tsv and csv
    • raw text data and word embeddings

Goals

We have a looong way to go :)

  • Fast
  • Easy to use
  • Portable
  • More control when you need it
  • Easy to define custom layers
  • Readable internal codebase

License

MIT