v5.0.0a1

Bugfix：

1. Mdarray compatibility support for snapshot operation

Features：

1. Refine batchnormalization and pooling interface for better performance and precise information
2. Enable LeakyRelu for intel64 backend

Limitation:

1. From this release, iDeep1.0.* will not be supported anymore, only iDeep2.0.0* is supported

v4.0.0b4

• Rebase to chainer v4.0.0b4, the master branch has been switched to chainer v4.0.0b4 development.
• Add poly learning rate update policy for training extension
• Enable lazy gradient sum
• Enable dilated convolution for arbitrary dx, dy

v4.0.0a1_i3.0.1a

Features:
	• Improve ideep4py interface to simplify integration.
	• Refactor ideep4py elementwise operation module.
	• Support tanh for mdarray.
	• Support a strategy of learning rate adjustment, poly.
	•  Resnet* with bs=128 performance  improved 20%
                Googlenetv2 Googlenetv3 improved 10%.
Bug fixings:
	•  Fix sum along axis bug in public format, which causes random crash.
	•  Fix MKLDNN build failures on ubuntu_16.04 gcc_5.4
Misc:
	• Add ubuntu Dockerfiles, benchmarks and training examples.

v4.0.0a1_i3.0.0a

1. Rebase to Chainer V4.0.0a1
2. Improve inference performance by eliminating limitation that reuse cc in Intel Chainer V2
3. Boost performance a lot on Inception V3(10x +) and SSD/VGG16(10x +)
4. Improve native framework to align APIs of ideep
5. Fix converge test issues: GoogLeNet V1 and VGG16 can converge to SOTA accuracy, Resnet50 can archive same accuracy as that of GPU

v2.0.0a_i2.0.5a

1. Fix converge test issues
2. GoogLeNet V1 can converge to SOTA accuracy.
3. Resnet50 can archive same accuracy as that of GPU.
4. VGG16 can converge to SOTA accuracy.

v2.0.0a_i2.0.4a

Inplace matrix element-wise mult
Speedup cosim comparing data process
Fix cosim bugs

v2.0.0a_i2.0.3a: Merge code and fix conflicts

• Element-wise Add/Subtract/Mult
• Fix cosim issue on concat layer
• Inplace memory for batch norm’s backward
• Improve setup system to prepare external library mkldnn automatically.
• Align result checking options with mkldnn
• Some bug fix in weight reorder, gy reuse, gpu tests, cosim test

i2.0.2a

Deconvolution acceleration on CPU.
Fix GPU path bugs.

i2.0.1a

Split mkl-dnn and Chainer.
Build enhancement. Update MKL-DNN mandatory.
Enhance condition checking for mkl-dnn fast path.
Python/MKL-DNN co-simulation enhancement.
Fix grad accumulate bug

i2.0.0a

Implement CPU acceleration for CNN layers based on MD-Array and Compute-Complex.