Tensorust is a general purpose tensor library written in Rust. It features no C/C++ bindings, which are otherwise commonly used in tensor libraries, e.g. Tensorflow and PyTorch. Tensorust is pure Rust. The heart of the library is dynamic building of a tensor DAG and performs automatic gradient calculation.
The current status of the library is in early v0.1 stages; experiments are being conducted to verify data layout in memory for optimal performance. The only datatype supported for the Tensor struct as of writing this is float32, but making it a generic type is work in progress.
The currently implemented Tensor operations are:
- add
Operations to implement are:
- sub
- mul
- dot
- matmul
- conv2d
- pool2d
- activation functions (tanh, relu, sigmoid, ...)
Below is an example showcasing how to create two Tensors and get a resulting Tensor from the add operation. Due to how the DAG construction works the resulting Tensor is going to be part of a context, meaning, it was created through a Function and thus have two parent Tensors
use tensorust::{Tensor};
// Create two Tensors of the same size; one with only zeros and
// the other with only ones. Add them and get resulting Tensor.
fn main() {
let a: Tensor = Tensor::zeros(vec![128, 64]);
let b: Tensor = Tensor::ones(vec![128, 64]);
let c: Tensor = a + &b;
// An alternative way to add the Tensors is to explicitly call
// let c: Tensor = a.add(&b);
}All code is written under an Apache-2.0 styled license, please see LICENSE for more information.