We are working on multiple new features and refactors (gluon, sparse, engine, etc) towards an 1.0 release. But there are also some legacy issues that needs to be resolved. Here is a list of issues I have noted. Feel free to raise new issues or contribute fixes.

@mli @tqchen @eric-haibin-lin @reminisce @asmushetzel @jermainewang @ptrendx

Basic

• Sort out int32, int64, index_t, mx_uint etc.
  Currently TShape uses int64_t but the front-end and back-end interface still use uint32_t for indices.
  This need cleaning up and int64_t interface need to be exposed through a new set of CAPI. The general policy going forward should be to use int64_t for indices/size and int32_t for number of dimensions. Signed int should be used for function arguments unless there is a really strong reason to use unsigned.
  Most indexing related interface should support negative indexing.
• Deprecate mshadow.
  Remove usage of mshadow template evaluations and replace with Kernel::Launch or hand written cpu/gpu kernels.
• Verify type and shape support for operators.
  Currently some operators don't support types other than fp32 for legacy reasons. Proper type support and/or documentation should be added.
  Currently some operators have limit support for tensor ranks (maximum 5 dims). The limit needs to be increased or removed if possible.
• move legacy operators to new nnvm registration.
  Conv, FC, BN etc should be refactored into stateless operators and use thread_local to store cudnn tensor descriptors.
• remove mkl experimental and use the new storage type interface.

Stretch goals

• Support pybind11 or cython interface for faster python API.
• Use variant type instead of string for operator argument parsing.