Skip to main content

Tensors

Tensors are the main data structure of TensorFlow.js, which EnergeticAI is based on.

What are tensors?​

They're a generalization of scalars, vectors and matrices to potentially higher dimensions. They have different ranks representing the number of dimensions that they have, and different shapes representing the size of each dimension.

TensorFlow.js provides a number of different tensor types:

  • Scalar: A tensor with rank 0, representing a single value.
  • Tensor1D: A tensor with rank 1, representing a vector of values.
  • Tensor2D: A tensor with rank 2, representing a matrix of values.
  • Tensor3D: A tensor with rank 3, representing a cube of values.
  • and so on...

Tensorflow also provides a handy guide about tensors here.

How do I create tensors?​

You can create tensors using the appropriate typed function, which takes values as input and returns a tensor as output:

import { scalar, tensor1d, tensor2d, tensor3d } from "@energetic-ai/core";

const scalarTensor = scalar(1);
const tensor1DTensor = tensor1d([1, 2, 3]);
const tensor2DTensor = tensor2d([
  [1, 2],
  [3, 4],
]);
const tensor3DTensor = tensor3d([
  [[1], [2]],
  [[3], [4]],
]);

What operations can you perform on them?​

You can perform a wide variety of operations on tensors, including:

  • Arithmetic operations: Addition, subtraction, multiplication, division, and more.
  • Matrix operations: Matrix multiplication, dot product, and more.
  • Reduction operations: Sum, mean, min, max, and more.
  • Convolution operations: Convolution, pooling, and more.

See the TensorFlow.js API docs for a full list of operations.

What if I just want an array or a number?​

If you've performed a set of operations on a Tensor and you want to get the result as a number or an array, you can use the array() or arraySync() methods:

import { tensor1d, dot, norm, div, mul } from "@energetic-ai/core";

function cosineSimilarity(embedding1: number[], embedding2: number[]): number {
  const tensor1 = tensor1d(embedding1);
  const tensor2 = tensor1d(embedding2);
  return div(
    dot(tensor1, tensor2),
    mul(norm(tensor1), norm(tensor2))
  ).arraySync() as number;
}

If a Tensor is rank 0, you can use arraySync() to get the value as a number. If a Tensor is rank 1 or above, you can use arraySync() to get the value as an array, potentially nested.