caffe.rs

A Rust FFI wrapper for the Caffe deep learning library, using rust-bindgen.

Setup

Requires a caffe distribution built with the patches in ajtulloch/caffe:caffe-ffi (https://github.com/ajtulloch/caffe/tree/caffe-ffi) to expose the necessary structures over FFI.

You can clone and build that repository as usual. Set the CAFFE_ROOT environment variable to allow the build.rs script to correctly generate dependencies.

Example

Inference on a pre-trained network

// Create the newtork
let mut net = caffe::Net::new(Path::new("test-data/lenet.prototxt"),
                              caffe::Phase::Test);
// Initialize the weights
net.copy_trained_layers_from(Path::new("test-data/lenet.caffemodel"));

// Fill in the input data blob.
let mut data_blob = net.blob("data");
let mut ones: Vec<_> = repeat(1.0 as f32)
                       .take(data_blob.len())
                       .collect();
data_blob.set_data(ones.as_mut_slice());

// Run a foward pass.
net.forward_prefilled();
let prob_blob = net.blob("prob");

// Process the output probabilities.
let probs = prob_blob.as_slice();
println!("{:?}", probs.to_vec());
assert_eq!(probs[0], 0.06494621)

Running a solver

let mut solver = caffe::Solver::new(
    Path::new("test-data/lenet_solver.prototxt"));
solver.solve();