* =caffe.rs=
A Rust FFI wrapper for the [[http://caffe.berkeleyvision.org/][Caffe]] deep learning library, using [[https://github.com/crabtw/rust-bindgen][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.

Note that to build and run the tests, you'll have to modify =src/ffi.rs= to
point to the right location for your installation. To get the tests passing,
update the dataset locations in =test-data/lenet_train_test.prototxt=.

** Example

*** Inference on a pre-trained network
#+BEGIN_SRC rust
  // 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)
#+END_SRC

*** Running a solver
#+BEGIN_SRC rust
  let mut solver = caffe::Solver::new(
      Path::new("test-data/lenet_solver.prototxt"));
  solver.solve();
#+END_SRC