Prepare for Keras-MXNet v2.2.2 release

[sandeep-krishnamurthy]

• Prepare for Keras-MXNet v2.2.2 release.
• Pull in Keras v2.2.2 and make required changes in Keras-MXNet.

Testing Done:

1. All UTs, ITs pass.
2. Smoke test on examples/ CNN and RNN usecases with channels_first and channels_last.

@kalyc @roywei

[sandeep-krishnamurthy]

Update on Testing:

Using Beta package pip install keras-mxnet --pre,

1. Tested all examples on AWS P3.8X instance with 1 GPU, multi-GPU setting.

[kalyc]

Thanks for your contributions @sandeep-krishnamurthy As discussed offline, I noticed that reuters_mlp_relu_vs_selu.py example doesn't work on CPU.

Here is the simple bind error message - InferShape pass cannot decide shapes for the following arguments (0s means unknown dimensions). Please consider providing them as inputs:

You mentioned that its because network2 doesn't support AlphaDropout So lets disable this example on our end?

Rest of the examples are working fine on CPU using the beta package. Tested on Mac & C5.18xlarge

[sandeep-krishnamurthy]

@kalyc - Thank you!