Align pow gradient with TF

[jgartman]

Description

Multiple problems with the gradient for pow were brought up in tensorflow/tfjs#346. One was fixed by #1376, the other is that the gradient with respect to the exponent is nan when the base is negative. This differs from TF which returns 0 in this case: /~https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/math_grad.py#L1046

---

For repository owners only:

Please remember to apply all applicable tags to your pull request.
Tags: FEATURE, BREAKING, BUG, PERF, DEV, DOC, SECURITY

For more info see: /~https://github.com/tensorflow/tfjs/blob/master/DEVELOPMENT.md

---

This change is 

[jgartman]

There was a test I added that was failing because of numerical precision issues and I decided to just remove it. The failure was because tf.greater is being used to compare the base of an exponent to 0 and on Safari when the base is 0 this comparison was returning true. I tried comparing the base to a small non-zero value like 1e-4 but the test still failed. Let me know if there are other suggestions, not sure how useful this PR is if it's not going to work as expected on all platforms.

[nsthorat]

Reviewable status: 0 of 1 approvals obtained (waiting on @jgartman)

---

src/ops/arithmetic_test.ts, line 820 at r1 (raw file):

  });

  it('gradient wrt exponent with non-positive base', () => {

negative :)

---

src/ops/binary_ops.ts, line 268 at r1 (raw file):

    };
    const derExp = () => {
      const condition = $base.greater(scalar(0));

no need to box the 0 in a scalar anymore

---

src/ops/binary_ops.ts, line 269 at r1 (raw file):

    const derExp = () => {
      const condition = $base.greater(scalar(0));
      const logBase = $base.log().toFloat().where(condition, zerosLike($base));

I think log will return a float.

also, where broadcasts you can just pass 0 directly (without boxing). This will reduce the memory footprint

[jgartman]

Reviewable status: 0 of 1 approvals obtained

---

src/ops/binary_ops.ts, line 269 at r1 (raw file):

Previously, nsthorat (Nikhil Thorat) wrote…

I think log will return a float.

also, where broadcasts you can just pass 0 directly (without boxing). This will reduce the memory footprint

Doesn't look like where broadcasts /~https://github.com/tensorflow/tfjs-core/blob/v0.14.2/src/ops/logical_ops.ts#L135-L165

[nsthorat]

Reviewed 2 of 2 files at r2.
Reviewable status: 0 of 1 approvals obtained

[nsthorat]

Hmm -- unfortunately the merge broke this :/ could you look into this if you have a sec?

[jgartman]

Just needed to update a recently added test to match the new behavior

[nsthorat]

Reviewed 1 of 1 files at r3.
Reviewable status: complete! 1 of 1 approvals obtained