New Features - Sparse Tensor Support

• Added comprehensive support for sparse matrices. See help on mx.sym.sparse and mx.nd.sparse for more info
  
• Added limited cpu support for two sparse formats in Symbol and NDArray - CSRNDArray and RowSparseNDArray
• Added a sparse dot product operator and many element-wise sparse operators
• Added a data iterator for sparse data input - LibSVMIter
• Added three optimizers for sparse gradient updates: Ftrl, SGD and Adam
• Added push and row_sparse_pull with RowSparseNDArray in distributed kvstore
  

New Features - Gluon and Autograd

• Added Gluon package, a high-level interface designed to be easy to use while keeping most of the flexibility of low level API. Gluon supports both imperative and symbolic programming, making it easy to train complex models imperatively with minimal impact on performance. Neural networks (and other machine learning models) can be defined and trained with gluon.nn and gluon.rnn packages. 

• gluon.Trainer now allows reading and setting learning rate with trainer.learning_rate property.

• Added autograd package, which enables automatic differentiation of NDArray operations. Enter code in a autograd.record() block to capture the exact path by which each variable was generated. Use *.backward() on any variable to backpropagate.

• Added mx.autograd.grad and experimental second order gradient support (most operators don't support second order gradient yet)

• Added new loss functions - SigmoidBinaryCrossEntropyLoss, CTCLoss, HuberLoss, HingeLoss, SquaredHingeLoss, LogisticLoss, TripletLoss

• Added ConvLSTM to gluon.contrib

• Autograd now supports cross-device graphs. Use x.copyto(mx.gpu(i)) and x.copyto(mx.cpu()) to do computation on multiple devices

Other New Features

• Added limited support for fancy indexing, which allows you to very quickly access and modify complicated subsets of an array's values. x[idx_arr0, idx_arr1, ..., idx_arrn] is now supported. Full support coming soon in next release. Checkout master to get a preview.
  

• Random number generators in mx.nd.random.* and mx.sym.random.* now supports both CPU and GPU

• NDArray and Symbol now supports "fluent" methods. You can now use x.exp() etc instead of mx.nd.exp(x) or mx.sym.exp(x)

• Added mx.rtc.CudaModule for writing and running CUDA kernels from python

• Added multi_precision option to optimizer for easier float16 training

Performance

• Enabled JIT compilation. Autograd and Gluon hybridize now use less memory and has faster speed. Performance is almost the same with old symbolic style code.
  
• Full support for NVidia Volta GPU Architecture and Cuda 9. Training is up to 3.5x faster than Pascal when using float16.

API Changes

• Operators like mx.sym.linalg_* and mx.sym.random_* are now moved to mx.sym.linalg.* and mx.sym.random.*. The old names are still available but deprecated.

• sample_* and random_* are now merged as random.*, which supports both scalar and  NDArray distribution parameters.

Bug-fixes

• Fixed a bug that causes argsort operator to fail on large tensors

• Fixed numerical stability issues when summing large tensors