autograd

Differentiable operations and tensors backed by ndarray.

Installation

[dependencies] autograd = { version = "0.9.8", features = ["mkl"] } mkl feature is recommended to speedup gemm operations using Intel MKL.

Features

Lazy, zero-copy tensor evaluation

Computation graphs are created on the fly (a.k.a define-by-run), but are not evaluated until Tensor::eval or ag::eval is called. This mechanism balances better performance and flexibility. ```rust extern crate autograd as ag;

let a: ag::Tensor = ag::ones(&[60]); let b: ag::Tensor = ag::ones(&[24]); let c: ag::Tensor = ag::reshape(a, &[3, 4, 5]); let d: ag::Tensor = ag::reshape(b, &[4, 3, 2]); let e: ag::Tensor = ag::tensordot(c, d, &[1, 0], &[0, 1]); e.eval(&[]); // Getting ndarray::Array here. ```

Reverse-mode automatic differentiation

There are a lot of built-in operations that support higher-order derivatives, and you can also define your own differentiable ops with ndarrays easily.

Here we are just computing partial derivatives of z = 2x^2 + 3y + 1.

```rust extern crate autograd as ag;

let ref x = ag::placeholder(&[]); let ref y = ag::placeholder(&[]); let ref z = 2.xx + 3.*y + 1.;

// dz/dy let gy = &ag::grad(&[z], &[y])[0]; println!("{:?}", gy.eval(&[])); // => Some(3.)

// dz/dx (requires to fill the placeholder x) let gx = &ag::grad(&[z], &[x])[0]; println!("{:?}", gx.eval(&[ag::Feed(x, ag::ndarray::arr0(2.).into_dyn().view())])); // => Some(8.)

// ddz/dx (differentiates z again) let ggx = &ag::grad(&[gx], &[x])[0]; println!("{:?}", ggx.eval(&[])); // => Some(4.) ```

Neural networks

This crate has various low-level features inspired by tensorflow/theano to train neural networks. ```rust // This is a softmax regression for MNIST digits classification with Adam. // This achieves 0.918 test accuracy after 3 epochs (0.11 sec/epoch on 2.7GHz Intel Core i5). let ref w = ag::variable(ag::ndarrayext::glorotuniform::(&[2828, 10])); let ref b = ag::variable(ag::ndarray_ext::zeros::(&[1, 10])); let ref x = ag::placeholder(&[-1, 2828]); let ref y = ag::placeholder(&[-1]); let ref z = ag::matmul(x, w) + b; let ref loss = ag::sparsesoftmaxcrossentropy(z, y); let ref params = [w, b]; let ref grads = ag::grad(&[loss], params); let ref predictions = ag::argmax(z, -1, true); let ref accuracy = ag::reducemean(&ag::equal(predictions, y), &[0], false); let ref adam = ag::gradientdescentops::Adam::default(); let mut statefulparams = ag::gradientdescentops::Adam::varswithstates(params); let ref updateops = adam.computeupdates(&statefulparams, grads);

// -- dataset -- let ((xtrain, ytrain), (xtest, ytest)) = dataset::load();

// -- training loop -- for epoch in 0..maxepoch { ... ag::eval(updateops, &[ag::Feed(x, xbatch), ag::Feed(y, ybatch)]); } ```

ConvNet, LSTM example can be found in examples

Hooks

You can register hooks on ag::Tensor objects for debugging. ```rust extern crate autograd as ag;

// .p() is a shorthand for .with(ag::Hook::Print). let a: ag::Tensor = ag::zeros(&[4, 2]).p(); let b: ag::Tensor = ag::ones(&[2, 3]); let c = ag::matmul(a, b);

c.eval(&[]); // Zeros: // [[0.0, 0.0], // [0.0, 0.0], // [0.0, 0.0], // [0.0, 0.0]] shape=[4, 2], strides=[2, 1], layout=C (0x1) ```

Why Rust?

• No need for bridges for fast languages. The entire logic including hotspots (kernels etc) is implemented in pure Rust, without compromising performance.

• Memory safety. For example, Rust's lifetime checker makes it possible to implement zero-copy computation graphs without GC.

For more, see documentation or examples