Prepare Keras-MXNet 2.2.4.1 release (#199)

* Prepare Keras-MXNet 2.2.4.1 release

* Fix typo

* Add documentation on sparse operators

Author: kalyc

docs/mxnet_backend/README.md

@@ -4,6 +4,6 @@
 2. [Multi GPU Model Training](multi_gpu_training.md)
 3. [Using RNN - Limitations and workarounds](using_rnn_with_mxnet_backend.md)
 4. [Save MXNet model from Keras-MXNet](save_mxnet_model.md)
-5. [Prepare Sparse Data for Training](prepare_sparse_data_for_training.md)
+5. [Using Sparse Data with Keras-MXNet](using_sparse_data_with_mxnet_backend.md)
 6. [Performance Tuning](performance_guide.md)
 7. [Benchmark Results](../../benchmark/README.md)

…kend/prepare_sparse_data_for_training.md …/using_sparse_data_with_mxnet_backend.mddocs/mxnet_backend/prepare_sparse_data_for_training.md renamed to docs/mxnet_backend/using_sparse_data_with_mxnet_backend.md docs/mxnet_backend/prepare_sparse_data_for_training.md renamed to docs/mxnet_backend/using_sparse_data_with_mxnet_backend.md

@@ -34,6 +34,18 @@ output = Dense(units=1, activation='linear', sparse_weight=True)(input)
 Using sparse weight is useful in models where we need to invoke 
 [dot operation](https://mxnet.incubator.apache.org/api/python/ndarray/sparse.html#mxnet.ndarray.sparse.dot) for training.
 In this code, setting sparse_weight to `True` will create a `row_sparse` tensor that will be used internally.
+### Sparse operators
+We have added sparse support for the following operators with MXNet backend:
+* sum
+* mean
+* dot
+* concat
+* Embedding
+For using Embedding layer with sparse data, we need to set flag `sparse_grad` to True
+```python
+embedding=Embedding(max_features, 128, input_length=10, sparse_grad=True)
+```
+Please see release notes for v2.2.4.1 for further details
 ## Slice sparse data
 As MXNet sparse arrays do not support reshape operation, it is essential to be able to slice the data for training the 
 model.

keras/__init__.py

@@ -23,4 +23,4 @@
 from .models import Model
 from .models import Sequential
 
-__version__ = '2.2.4'
+__version__ = '2.2.4.1'

keras/backend/mxnet_backend.py

@@ -2786,7 +2786,7 @@ def rnn(step_function, inputs, initial_states,
         # TODO: remove version check after mxnet 1.3.1 release
         if mx.__version__ < '1.3.1':
             raise NotImplementedError('unroll=False in RNN only works with MXNet 1.3.1 or newer, '
-                                      'please upgrade to latest master using: pip install --ugprade mxnet --pre')
+                                      'please upgrade to latest master using: pip install --upgrade mxnet --pre')
         # defining step functions for each RNN cells, implementation taken from call functions
         # from each RNN cell class in keras.layers.recurrent
 

setup.py

@@ -26,7 +26,7 @@
 '''
 
 setup(name='keras-mxnet',
-      version='2.2.4',
+      version='2.2.4.1',
       description='Deep Learning for humans. Keras with highly scalable,\
                    high performance Apache MXNet backend support.',
       long_description=long_description,