v2.0.0

 Bug fix:

1. Fix buffer issue in tensor
2. Fix snapshot resume bug
3. Remove useless libmklml_gnu.so file
4. Fix PYPI package building sequence issue

 New Feature:

1. Support channel shuffle operator and refine fast_math to support gcc 4.8-

 Others:

1. Add Docker and Conda files

v2.0.0_b1

1.Improve iDeep python package ReLU API to support Leaky-ReLU.
2.Improve iDeep python package BatchNormalization and Pooling2D API for better performance and precision.
3.Avoid AVX-SSE transition penalties to improve NMT performance about xxx%.
4.Bugfix for AlexNet training divergence.
5.Bugfix for AlexNet/VGG/SSD 3-5 percentage drop from SOTA.
6.Improve iDeep tensor(mdarray) compatibility.

v2.0.0_pr1: Update README

• Implement simple and general stateless C++ API to improve programmable.
• Implement CNN primitive computation LRU, scratch allocator and self-adapted data format optimization to maximize MKLDNN performance.
• Integrate iDeep python package (ideep4py) with stateless C++ API.

v1.0.4

• Update conda and docker files
• Improve iDeep MDarray compatibility about array reshape & python buffer protocol.
• Update mkldnn version to '464c268e544bae26f9b85a2acb9122c766a4c396'
• Restrict upper limit of build process to the number of CPUs(thanks@tkng)

v1.0.3

• Remove glog dependence
• Add docker and Conda files, update PYPI packages
• Improve ideep enabling condition to avoid potential performance drop in backward.
• Fix bug in split_axis
• Add interface of multi_add
• Replace omp parallel with omp simd for mlsl

v1.0.2

• Update test cases.
• Specify NumPy dependency version (NumPy 1.13 for now).
• Fix default parameter bug in sum along axis
• Fix bug in concat and split axis.

Update platform tag and version (v1.0.1)

• Support Pypi.
• Fix algorithm bug in sum along axis with public format.
• Enable Travis test for ideep github.

ideep v1.0.0

First release of ideep python package ideep4py.