• TensorJS
• How to use
  • Tensors
    • Tensor operations
    • Reading values
    • Data types
    • Converting between backends
  • Onnx model support
    • Optimizations
    • Running with half precision
    • Other performance considerations
  • Autograd functionality
  • Sparse tensors
    • Backend support for sparse tensors
• Documentation
• Contributing
• Development

This is a JS/TS library for accelerated tensor computation intended to be run in the browser. It contains an implementation for numpy-style multidimensional arrays and their operators.

It also allows executing Onnx models. For examples check the examples folder.

There are three execution backends available:

• CPU: This is implemented in plain javascript and thus not very fast. It is intended to be a reference implementation. Big optimizations are avoided for simplicity.
• Web Assembly: This is implemented in Rust. It is optimized for faster execution (although right now there is a lot of work to be done).
• GPU: This uses WebGL to enable very fast execution and should be used whenever a GPU is available. It is typically ~10-100 times faster than the WASM backend (except for a few operators). Most of the development focus was spent here so this is by far the fastest backend.

Install with

$ npm install @hoff97/tensor-js

and then import

import * as tjs from '@hoff97/tensor-js';

or import the stuff you need directly.

You can create tensors of the respective backend like this:

• CPU:

  const tensor = new tjs.tensor.cpu.CPUTensor([2,2], [1,2,3,4]);

• WASM:

  const tensor = new tjs.tensor.wasm.WASMTensor(new Float32Array([1,2,3,4]), [2,2]);

• GPU:

  const tensor = new tjs.tensor.gpu.GPUTensor(new Float32Array([1,2,3,4]), [2,2]);

  or directly from an image/video element:

  const video: HTMLVideoElement = document.querySelector("#videoElement");
  const tensor = tjs.tensor.gpu.GPUTensor.fromData(video);

  which will be a tensor with shape [height,width,4] and data type float32. Creating a GPU tensor from a video element will usually be pretty fast. Creation from an image not necessarily, since here the image data first has to be transferred to the GPU.

Once you have created a tensor, you can do operations on it, for example:

• Add two tensors

  const res = a.add(b);

• Matrix multiplication

  const res = a.matMul(b);

• Find the maximum

  const res = a.max(1);

For a list of all operators, see the docs. Most operators will behave like their numpy/pytorch counterparts.

When you want to read data from a tensor:

const values = await tensor.getValues();

which will give you the values as a array of the values. For CPU tensors you can also get the value at an index:

const value = tensor.get([1,2,3,4]);

Tensors are created with float values (using 32 bits) by default. You can specify another data type on creation:

  const tensor = new tjs.tensor.cpu.CPUTensor([2,2], [1,2,3,4], 'float16');

or cast to another data type with:

  const casted = tensor.cast('float16');

The available data types are float64, float32, float16, int32, int16, int8, uint32, uint16, uint8. Note that not all backends support all data types:

• CPU: Supports all data types, but float16 will be represented as float32 internally
• WASM: Supports all except float16
• GPU: Supports all except float64. Note that except for float16, all other data types will be represented by float32 internally, since WebGL1 does not allow writing anything else than floats to frame buffers. This means that for int32 and uint32, not the whole range of values of the respective data type is available.

The data type of a tensor can be accessed via tensor.dtype. Additionally, each tensor has a generic type argument, which will carry its data type:

  const tensor: Tensor<'float16'> = new tjs.tensor.cpu.CPUTensor([2,2], [1,2,3,4], 'float16');

This allows type checking tensor operations, which means that only tensor operations with the same data type compile when using typescript. The generic type defaults to float32. If you want to represent the data type of a tensor with an unknown data type, write for example

  const tensor: Tensor<any> = a.add(b);

or alternatively

  const tensor: Tensor<DType> = a.add(b);

You can conver a tensor to a different backend like so:

const cpuTensor = await tjs.util.convert.toCPU(tensor);
const wasmTensor = await tjs.util.convert.toWASM(tensor);
const gpuTensor = await tjs.util.convert.toGPU(tensor);

Note that converting between backends (especially from/to WebGL) is an expensive operation and should be prevented if possible!

You can load an onnx model like this:

const respones = await fetch(`model.onnx`);
const buffer = await res.arrayBuffer();

const model = new tjs.onnx.model.OnnxModel(buffer);

To see all supported operators, check the supported operator list.

You will very likely want to run this model on the GPU. To do this:

await model.toGPU();

There are a few optimization passes that can be done on an Onnx model to get faster execution. To do this, run

model.optimize()

By default full precision floats (32-bits) are used for model execution. On the GPU backend, you can try executing with half precision, but be aware that this might not work for all models. To use half precision, specify this when loading the model:

const model = new tjs.onnx.model.OnnxModel(buffer, {
  precision: 16
})
model.toGPU();

For the best performance you should also create your GPU tensors with half precision

const tensor = new tjs.tensor.gpu.GPUTensor(new Float32Array([1,2,3,4]), [2,2], 'float16');

The outputs of the model will be half-precision tensors as well. To read the values of a half precision gpu tensor, you have to convert it to full precision first, which can be done with:

const values = await tensor.cast('float32').getValues();

Try to run your models with static input sizes. TensorJS will compile specialized versions of all operations after enough forward passes. For this the input shapes of the tensors have to be constant though.

Automatic differentiation is supported. For this create variables from all your tensors:

const a = new tjs.tensor.cpu.CPUTensor([2,2], [1,2,3,4]);
const b = new tjs.tensor.cpu.CPUTensor([2,2], [5,6,7,8]);

const varA = new tjs.autograd.Variable(a);
const varB = new tjs.autograd.Variable(b);

Or use the utility methods:

const varA = tjs.autograd.Variable.create([2,2], [1,2,3,4], 'GPU');
const videoElement = document.querySelector("#videoElement");
const varB = tjs.autograd.Variable.fromData(videoElement);

Afterwards you can perform normal tensor operations:

const mul = varA.matMul(varB);
const sum = mul.sum();

To perform a backward pass, call backward on a scalar tensor (a tensor with shape [1]). All variables will have an attribute .grad, which is the gradient

sum.backward();

console.log(varA.grad);

Multiple backward passes will add up the gradients. After you are done with the variable, delete the computation graph by calling delete().

Sparse tensors are tensors where most entries are zero, for example the following one:

const a = new CPUTensor([3,3],
  [1,0,0,
   0,2,0,
   0,3,4]);

TensorJS supports sparse tensors in coordinate format, where we store the coordinates and values of the nonzero entries in two tensors:

  const indices = [
    0,0,  // Corresponds to value 1
    1,1,  // Corresponds to value 2
    2,1,  // Corresponds to value 3
    2,2   // Corresponds to value 4
  ];
  const indiceTensor = new CPUTensor([4, 2], indices, 'uint32');

  const values = [1,2,3,4];
  const valueTensor = new CPUTensor([4],values);
  const sparseTensor = new SparseTensor(valueTensor, indiceTensor, [3,3]);

The implementations of the operators for sparse tensors only consider the nonzero entries and are thus faster than their dense counterparts.

Note that some operators make specific assumptions on the sparse tensor, for details check the corresponding documentation here.

As of now, most operators are only supported on the CPU and WASM backend. If an operation is not supported, this is noted in the docs.

You can find the documentation here.

See Contributing.md

See Development.md