Tensorflow.js lab

Setup

//Clone the repo:
git clone https://github.com/kaueDM/tensorflow-lab

//Install dependencies:
npm install

//If you dont have Parcel installed:
npm install -g parcel-bundler

//Run Parcel server:
parcel index.html

Tensors

Tensor: The main building block of a Tensorflow neural net. Its a generic name for any N-dimensional thing of numbers and it's operations.

Examples of 'thing of numbers':

Scalar: A single number. Just it:

4. Or maybe 5.

Vector: Fancy word for Array:

[1,2,3,4, ...n]

Matrix: A 2-dimensional grid:

//I have no idea how i can represent this, so, this is my best:
| 1 2 3 |
| 4 5 6 |
| 7 8 9 |

Building a Tensor:

tf.tensor(values, shape?, dtype?)

values: Kinda obvious, huh?

shape: Dimension of a Tensor.

It's 2x2 matrix? Easy, [2,2]

Or 3x1 matrix populated with 2x2 matrices? Not a problem: [3,2,2]

dtype: DataType. float32, int32 or boolean

Example

Creating a Tensor with 5x3 (Rows x Columns, always) shape using 15 random numbers between 1 and 100

Input

const values = Array.from({ length: 15 }, _ => Math.floor(Math.random() * 100) + 1)
const shape = [5,3]
tf.tensor(values, shape).print()

Output

[
    [29, 39, 42],
    [35, 11, 79],
    [16, 49, 82],
    [46, 66, 29],
    [53, 47, 76]
]

You understand how to create a Tensor now? Great! Now forget it and use the proper method for each type of tensor. Doing this, you get a more readable code.

• tf.scalar (value, dtype?): Rank-0 tensor (scalar). value is a single number.
• tf.tensor1d (values, dtype?): Rank-1 tensor. values is an array of numbers.
• tf.tensor2d (values, shape?, dtype?): Rank-2 tensor. values is a matrix of numbers.
• tf.tensor2d (values, shape?, dtype?): Rank-3 tensor. values is a matrix of matrices. Exactly what we got in the example above.

Operations

Matrices multiplication

tf.matMul (a, b, transposeA?, transposeB?)

• Works with rank-2 tensors only.
• The number of cols in a should match the rows in b. Or TRANSPOSE one of them.
• To transpose a rank-2 tensor, you should assign the new tensor (Tensors are IMMUTABLE!) to a new variable, or you will get a big error.

Memory management

First of all, Garbage collector is dead in TF.js lands. Why? Because we are working in the GPU. Tf.js needs to manage 'manually' the GPU memory to achieve a good speed in mathematical operations. So, pay attention to memory leaks.

How much tensors do i have?

tf.memory().numTensors will give you the answer.

After you used your tensors, you can manually dispose them using dispose() method.

const values = Array.from({ length: 6 }, _ => Math.floor(Math.random() * 6) + 1)
const shape = [2, 3]
const myFirstTensor = tf.tensor2d(values, shape)
const mySecondTensor = tf.tensor2d(values, shape)

//Do something meaningful with your tensor here...
const transposedTensor = mySecondTensor.transpose()
const multipliedMatrices = myFirstTensor.matMul(transposedTensor)

multipliedMatrices.print()

//...and dispose when they aren't necessary anymore
myFirstTensor.dispose()
mySecondTensor.dispose()
transposedTensor.dispose()

OR

Wrap your code with tf.tidy() and let the memory cleanup happens automagically:

tf.tidy(_ => {
    const values = Array.from({ length: 6 }, _ => Math.floor(Math.random() * 6) + 1)
    const shape = [2, 3]
    const myFirstTensor = tf.tensor2d(values, shape)
    const mySecondTensor = tf.tensor2d(values, shape)
    const transposedTensor = mySecondTensor.transpose()
    const multipliedMatrices = myFirstTensor.matMul(transposedTensor)

    multipliedMatrices.print()
})

If you don't want to dispose a tensor automatically inside the tf.tidy() method, you can call tf.keep() to, guess what, keep it.

tf.tidy(_ => {
    const values = Array.from({ length: 6 }, _ => Math.floor(Math.random() * 6) + 1)
    const shape = [2, 3]
    const myFirstTensor = tf.tensor2d(values, shape)
    const mySecondTensor = tf.tensor2d(values, shape)
    
    tf.keep(myFirstTensor)
    //Only mySecondTensor will be disposed.
})