tensorflow: Tensor.__bool__'s signature is incorrect for one element tensors

[hoel-bagard]

CurrentlyTensor.__bool__'s signature is as follows:

class Tensor:
    def __bool__(self) -> NoReturn: ...

It suggests that the method should never be called, however the following is valid code:

import tensorflow as tf

bool(tf.constant([1]))

Since this depends on the shape of the tensor, I don't know if this can be fixed, but it would be nice to have.

[srittau]

Cc @hmc-cs-mdrissi who originally contributed the tensorflow stubs.

[hmc-cs-mdrissi]

There's two reasons for this.

1. It only works for scalars and fails with other shapes.
2. It only works in some execution modes. It'll be fine in eager, but can crash for graph mode even with scalars. Most code I work with needs to be graph mode compatible.

I'm fine with the change. My experience leans that those two issues cover large percent of usages, but it's fair that policy is prefer false negatives and I don't have actual numbers for it.