Tfjs: Tensor/Matrix and math operations + reduction

[JasonShin]

To get help from the community, check out our Google group.

TensorFlow.js version

0.13.2

Describe the problem or feature request

I'd like to request a feature reduce that cumulatively applies the math function onto each item in the Tensor and produce a new tensor with a new shape.

Such API might be invoked like

tf.tensor2d([[1, 2], [3, 4]]).add.reduce(0)
// column wise reduction
// Result: [4, 6]<tensor>

tf.tensor2d([[1, 2], [3, 4]]).add.reduce(1)
// Row wise reduction
// Result: [3, 7]<tensor>

This feature request is heavily influenced by https://docs.scipy.org/doc/numpy-1.15.0/reference/generated/numpy.ufunc.reduce.html

[devjiro76]

This is alway i think, I totally agree with Jason

[dsmilkov]

TF.js API is closely aligned with the TF Python API. Does such a reduce feature exist in TF Python?

[JasonShin]

I couldn't find a TF Python API that is close to reduce ;(

[manrajgrover]

@JasonShin This should work:

const x = tf.tensor2d([[1, 2], [3, 4]]);

x.sum(axis=1).print(); // [3, 7]

You can read more about reduce operations here

[JasonShin]

@manrajgrover We were actually looking for both prod and sum methods but they weren't showing up on https://js.tensorflow.org/api/0.13.0/, which is the default doc site that the Google search result takes you to!

[manrajgrover]

Ah, yes it does look like prod reduction op was recently added. Seems like this issue can be closed now!

/cc @dsmilkov