nickmcintyre / numero Goto Github PK

The function createTensor() and tensor instance methods seem to be working fine in global mode, but no luck with static methods. The following statement throws an error.

const domain = Tensor.range(0, width);
// ^ throws TypeError: Tensor.range is not a function

Prefixing the static method call with p5.prototype does work. Hmm.

Tensor operations should work with p5.Vector where possible. p5.Vector could be extended to operate on Tensors.

I propose starting with a few test cases for matrix-vector operations.

NumPy's universal functions are nice and have analogs in TensorFlow.js. A core set of functions like num.add(t1, t2) would be useful.

Forcing beginners to use num.tidy() out of the gate steepens the learning curve. Setting the default backend to cpu would simplify this:

const b = num.tidy(() => {
    const a = createTensor([[1, 2], [3, 4]]);
    const x = createTensor([5, 6]);
    return a.dot(x);
});

print(b.toString());

to this:

const a = createTensor([[1, 2], [3, 4]]);
const x = createTensor([5, 6]);
const b = a.dot(x);

print(b.toString());

Let's 😉

Matrices (rank-2 tensors) should support elementary row operations. It'd be neat to use them to teach elimination.

It'd be nice to expose a manageable subset of tfjs' tensor methods.

Creating

Static methods for creating tensors likezeros(), ones(), eye() (for matrices), fill(), range(), linspace(), random() (mapped to tfc.randomUniform()), and randomGaussian() (mapped to tfc.randomNormal()) to begin. copy() (mapped to tfc.clone()) should be an instance method.

Transformations

Transformations should include reshape() and flatten(). Static versions should return a new tensor.

Slicing and Joining

concat(), reverse(), slice(), split(), stack(), and unstack() should work with p5.Vectors where possible. Static versions should return a new tensor.

Operations

Let's start with the basic math that overlaps with core p5.js math functions. Trigonometry should probably respond to angleMode(). Static versions should return a new tensor.

Symbolic expressions should map to compatible functions/objects where possible. SymPy's printing system is a good example of how this might work.