Is there a way to figure out if an NDArray is trainable/has a gradient?

[theRoughCode]

Is there a way to figure out if an NDArray variable has a gradient/requires backprop, similar to the requires_grad attribute in PyTorch?

I've tried checking whether x.grad is None, but that doesn't work for intermediate variables:

import mxnet as mx

x = mx.nd.array([1, 2])
x.attach_grad()
with mx.autograd.record():
  y = x**2
  z = 2 * y
z.backward()

>>> print(x.grad)
[4. 8.]
<NDArray 2 @cpu(0)>
>>> print(y.grad)
None

In this example, x has a gradient but the intermediate variable y does not, but it should also be trainable.

I've tried looking through the source code, but can't figure it out.

Corresponding SO post: https://stackoverflow.com/questions/64473958/is-there-a-way-to-figure-out-if-an-ndarray-is-trainable

[szha]

There isn't a direct way to access this information at the moment. It can be found in the backward operators as the req variable but it's not easy to tie this back to an NDArray.

What do you need this information for?

[theRoughCode]

Hey @szha, thanks for the response! I'm working on a Quantum Machine Learning library called PennyLane and we're planning on adding MXNet support as one of our ML backends. Specifically, we need to figure out which tensors are differentiable in order to perform automatic differentiation and avoid unnecessary computation for non-trainable tensors.

Are there any plans to add support for accessing this information (i.e. with an attribute similar to PyTorch)?