Crate coaster_nn[−][src]

Expand description

Provides a Coaster Plugin, to extend Coaster with Neural Network related operations such as convolutions, pooling, ReLU, etc. A full list of operations provided by this Plugin, can be found at the provided Operations section.

Overview

This Coaster Plugin extends Coaster’s Backend with NN related methods/operations. This allows you to run, these operations (and therefore your application) on your local machine as well as on servers, mobiles or any other machine (as if they were written for common CPU execution), while receiving the significant performance increases (usually one-to-two orders of magnitutde), by executing the operations on special purpose hardware such as GPUs - if they are available. Usage examples can be found in the next section.

The architecture of a Plugin is quite easy. It defines one Plugin Trait, in this case the NN trait, which implements basic functionality for initialization and multiple Plugin Operation Traits which define the methods which are going to be available on the Backed, as the Plugin Trait as well as the Plugin Operations Traits are implemented for the Coaster Backends (CUDA, OpenCL, Native). The operations take as arguments one or many SharedTensors, holding the data over which the operation should happen, and none or one Operation Configuration.

Usage

An example on how to write some data into a SharedTensor and compute the result of the sigmoid function for each value:

extern crate coaster as co;
extern crate coaster_nn as nn;
use co::prelude::*;
use co::frameworks::native::flatbox::FlatBox;
use nn::*;

fn write_to_memory<T: Copy>(mem: &mut FlatBox, data: &[T]) {
    let mut mem_buffer = mem.as_mut_slice::<T>();
    for (index, datum) in data.iter().enumerate() {
        mem_buffer[index] = *datum;
    }
}

use crate::co::frameworks::cuda::get_cuda_backend;
pub fn main() {
    // Initialize a CUDA Backend.
    // Usually you would not use CUDA but let Coaster pick what is available on the machine.
    let backend = get_cuda_backend();

    // Initialize two SharedTensors.
    let mut x = SharedTensor::<f32>::new(&(1, 1, 3));
    let mut result = SharedTensor::<f32>::new(&(1, 1, 3));
    // Fill `x` with some data.
    let payload: &[f32] = &::std::iter::repeat(1f32).take(x.capacity()).collect::<Vec<f32>>();
    let native = Native::new();
    let cpu = native.new_device(native.hardwares()).unwrap();
    write_to_memory(x.write_only(&cpu).unwrap(), payload); // Write to native host memory.
    // Run the sigmoid operation, provided by the NN Plugin, on your CUDA enabled GPU.
    backend.sigmoid(&mut x, &mut result).unwrap();
    // See the result.
    println!("{:?}", result.read(&cpu).unwrap().as_slice::<f64>());
}

Provided Operations

This Plugins provides the following operations. If not denoted otherwise, this means forward and backward A - means not yet implemented.

Operation	CUDA	OpenCL	Native
Sigmoid	cuDNN v5 or later	-	Rust
SigmoidPointwise	cuDNN v5 or later	-	Rust
ReLU	cuDNN v5 or later	-	Rust
ReLUPointwise	cuDNN v5 or later	-	Rust
Tanh	cuDNN v5 or later	-	Rust
TanhPointwise	cuDNN v5 or later	-	Rust

Normalization (LRN)	cuDNN v5 or later	-	-

Dropout	cuDNN v5 or later	-	Rust

Convolution	cuDNN v5 or later	-	Rust(fwd)

Softmax	cuDNN v5 or later	-	Rust
LogSoftmax	cuDNN v5 or later	-	Rust

Pooling Max	cuDNN v5 or later	-	Rust(fwd)
Pooling Avg	cuDNN v5 or later	-	-

Modules

frameworks

Provides the specific Framework implementations for the Library Operations.

Macros

impl_ops_log_softmax_for

log softmax impl generation macro

impl_ops_lrn_for

lrn impl generation macro TODO it’s all unimplemented!() right now

impl_ops_relu_for

relu impl generation macro

impl_ops_sigmoid_for

sigmoid impl generation macro

impl_ops_softmax_for

softmax impl generation macro

impl_ops_tanh_for

tanh impl generation macro

Enums

ConvBackwardDataAlgo

Different algorithms to compute the gradient with respect to the filter.

ConvBackwardFilterAlgo

Different algorithms to compute the gradient with respect to the filter.

ConvForwardAlgo

Different algorithms to compute the convolution forward algorithm.

DirectionMode

Direction Mode for RNN [cudnnDirectionMode_t][1] [1]: https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnDirectionMode_t

MathType

Indicate if Tensor Core Operations are permitted [cudnnMathType_t][1] [1]: https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnMathType_t

RnnAlgorithm

Algorithm for RNN [cudnnRNNAlgo_t][1] [1]: https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnRNNAlgo_t

RnnInputMode

Input Modes for RNN [cudnnRNNInputMode_t][1] [1]: https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnRNNInputMode_t

RnnNetworkMode

Network Type for RNN Networks [cudnnRNNMOde_t][1] [1]: https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnRNNMode_t

RnnPaddingMode

Enables/Disables the padded input/output [cudnnRNNPaddingMode_t][1] [1]: https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnRNNPaddingMode_t