Alpha release for Keras with MXNet backend.

@howard0su

Alpha

Sparse operations are not supported - ctc_label_dense_to_sparse, ctc_batch_cost, ctc_decode.
Few backend utility functions - stack, gradients, switch.
separable_conv2d - 2D convolution with separable filters are not supported
random_binomial is not supported.
Higher order utility functions - map, foldl, foldr.
Cannot use Numpy’s COOrdinate representation of the array: No sparse support.
Eigenvalue regularizer is not supported.
With MXNet backend, Model concatenation is not supported.

Batch Normalization does not support mode 0 or 2 with MXNet backend.
Optimizer states are not preserved when saving the model, with MXNet backend.
If the batch size is too small and you are using multiple GPUs for training, MXNet backend fails to split the data gracefully and errors out.
SoftmaxCrossEntropy only accepts 1D labels.
MXNet Categorical_crossentropy doesn't support from_logits.
Few failing unit tests and cause for failure - #65
Deconvolution operator with large target shape fails with MXNet backend.

Pip installation for dmlc/keras with MXNet backend. Will be added in Beta release.
Latest MXNet v0.10 has few missing functionalities to work with Keras. You need to build from latest source. 3. Refer installation guide in the next section below. Pip wheel will be made available for Beta release.
Performance improvements are not focussed in this release. MXNet’s full power may not be harnessed.
Training a model with other backends (ex: TF) and loading it with MXNet backend is not tested. This functionality is not supported.
Needs more code documentation for MXNet specific code.
Needs code style fix and clean up.

Thanks to all contributors who made significant contributions to get MXNet backend for Keras - @howard0su, @yajiedesign, @piiswrong, @wayao