distort_image_ops_test.py fails on tf-nightly-2.0-gpu

[WindQAQ]

The CI Build for PY2-GPU failed because of distort_image_ops_test.py.
https://source.cloud.google.com/results/invocations/7ed10b8d-6cba-47d9-b764-d823365528f4/log

Ideally, the manager should catch the exception raised in CC file

addons/tensorflow_addons/image/distort_image_ops_test.py

Lines 110 to 112 in 436580c

	with self.assertRaisesOpError("input must have 3 channels "
	"but instead has 4 channels"):
	self.evaluate(self._adjust_hue_in_yiq_tf(x_np, delta_h))

addons/tensorflow_addons/custom_ops/image/cc/kernels/adjust_hsv_in_yiq_op.cc

Lines 70 to 73 in 436580c

	OP_REQUIRES(
	context, channels == kChannelSize,
	errors::InvalidArgument("input must have 3 channels but instead has ",
	channels, " channels."));

[seanpmorgan]

So more info:
Seems not related to py2 vs py3. There has not been a gpu-py34 released since 20190718 so that's why it's passing.

The error is incredibly cryptic when decorated with tf.function... and when I remove that it looks like 2 errors are being raised instead of the one we're catching:

InvalidArgumentError: input must be at least 3-D, got shape[2,3] [Op:AdjustHsvInYiq]

and

ValueError: Shape must be at least rank 3 but is rank 2 for 'adjust_hsv_in_yiq/AdjustHsvInYiq' (op: 'AdjustHsvInYiq') with input shapes: [2,3], [], [], [].

I have no idea which commit started this and no reason why it's only happening on the GPU pip package.

[feihugis]

It can run in the eager mode but failed in the graph mode. I tried to add the GPU kernel but the build failed. Will update it here if I get any progress.

[seanpmorgan]

For tracking purposes... a temporary fix was put in #378