Add support for tensors with 0 in shape

[dsmilkov]

Fixes tensorflow/tfjs#567 (see the issue for details)

FEATURE

---

This change is 

[dsmilkov]

Let me know if you have ideas for other ops to test (in addition to matmul, add and concat). Asking since as you can see, the fix sometimes is op-specific, like in the concat case.

[pyu10055]

LGTM, you might want to check logic op for example two zero size tensors x > y, expandDim seems to be special for zero size tensors.

Reviewed 11 of 11 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov, @nsthorat, @tafsiri, and @pyu10055)

---

src/util.ts, line 265 at r1 (raw file):

      }
      implicitIdx = i;
    } else if (shape[i] <= 0) {

shape[i] < 0 ?

---

src/ops/broadcast_util.ts, line 87 at r1 (raw file):

      b = 1;
    }
    if (a === 1) {

does this method support broadcast [0,2], [2,2] => [0,2] ?

[nsthorat]

Reviewed 11 of 11 files at r1.
Reviewable status: complete! 1 of 1 approvals obtained (waiting on @dsmilkov, @nsthorat, and @tafsiri)

[dsmilkov]

Thanks. Added unit tests for expandDim, and tf.greater to verify it works with 0-sized tensors.

Reviewable status: complete! 2 of 1 approvals obtained (waiting on @dsmilkov and @tafsiri)

---

src/util.ts, line 265 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

shape[i] < 0 ?

Ah nice find, done. The first if fires for shape >= 0, so this if would never fire for 0, but good to be precise.

---

src/ops/broadcast_util.ts, line 87 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

does this method support broadcast [0,2], [2,2] => [0,2] ?

Yes, see the added unit tests in broadcast_util_test.ts.