flashlight_tensor

Tensor library written in pure rust, designed mostly for matrix operations

project not related to similarly named flashlight. The name was coincidental and chosen independently.

Features

• n-dimensional tensors
• Element-wise operations
• Scalar multiplication and addition
• Tensor multiplication and addition
• Matrix transformation
• ReLU and sigmoid activations
• forward/backward propagation operations on gpu
• CPU and GPU support
• GpuRunner
• Chunking

Instalation

[dependencies]
flashlight_tensor = "0.4.5"

# Experimental(Not everything documented and working. Use at your own risk)
flashlight_tensor = { git = "https://github.com/Bejmach/flashlight_tensor"}

Documentation

Docs

Quick Start

For gpu usage go to examples on github

use flashlight_tensor::prelude::*;

fn main(){
    //2 rows, 2 collumns, fill with 1.0
    let a: Tensor<f32> = Tensor::fill(1.0, &[2, 2]);
}

Tests

Run tests with:
cargo test

Patch notes

V0.2.4:

• matrix_vec/col, now return a matrix, not vector
• matrix_col/row_sum/prod, return a sum/product of all collumns/rows in matrix

V0.2.5

• Propably something

V0.2.6

• better file structure
• mutable operations for iterative functions

V0.2.6

• activation functions for neural network

V0.3.0

• gpu operations + docs

V0.3.1

• gpu only backward/forward propagation merged operations, kinda hard to perform, will try to abstact it into gpu_runner
• examples, with merged machine learning operations runtime

V0.4.0

• gpu_chunking
• gpu_runner

V0.4.1

• more tests
• better docs
• no need to tell output size in sample

V0.4.2

• gpu ml crucial bug fixes(seriously. earlier versions were unusable for gpu ml)
• removed merged backprop
• better tests

V0.4.3

• better file structure
• ml derivative ops that I forgot to include before

V0.4.4

• bug fixes
• backward activations

V0.4.5

• bug fixes

What changed in 0.4.x

• less code for similar result

Old way

let mut gpu_data = GpuData::new();
gpu_data.disable_shapes();

let sample = Sample::from_data(vec!{Tensor::fill(1.0, &[2, 2])}, vec!{1.0}, &[2, 2]);
gpu_data.append(sample);

let mut buffers = GpuBuffers::init(1, MemoryMetric::GB, &mut gpu_data, 0).await;
buffers.set_shader(&GpuOperations::Add);
buffers.prepare();

let full_gpu_output: Vec<Tensor<f32>> = buffers.run().await;

New way

New way also has integrated chunking, so if data you try to process it bigger than max buffer size, then it will run the operation in chunks and merge data at the end

You also dont need to set output size for sample. GpuRunner handles that

let mut runner: GpuRunner = GpuRunner::init(1, MemoryMetric::GB);

let sample = Sample::from_data(vec!{Tensor::fill(1.0, &[2, 2])}, vec!{1.0}, &[]);
runner.append(sample);

let output_data: Vec<Tensor<f32>> = runner.add().await;

Plans for 0.5.0

• nothing for now