* add retry on mac

* fix some error

* fix with some errors

* add failed unittests retry on win system

* modified the value of retry times

* fix the localsgd

* change configs to config & update examples

* fix deprecate decorator conflict

* add auto parallel L1 implementation
test=develop

fix the bug for dtype check for the argmin/argmax

test=develop

* fix dimensions error for mobilenetv1_KL_quant

fixes AssertionError: The size of weight scales vector (1000) does not match the number of output channels (1024) in the weights tensor fc7_weights.

add mul test

* remove comment

* add third case unit test

remove the dependence of matplotlib  in dataset of uci_housing

* Modify data download function, and support unittests of inference APIs on windows, test=develop

* The import error compatible with py2 and py3, and fix unittests problems of inference APIs on Windows, test=develop

* add lstm cudnn of padding data and refine cudnn codes

* add descriptor cache

* fix self arugments

* deal case if instance is None

* clean code

* fix usage

* add TranslatedLayer.program method

* add unittests & update example code

* polish example code

* update program api example code

update the check for the dtype check for the argmin, argmax

* - Added fix to mobilenet

* - compilation fix

* - Fix to conv2d grad oneDNN with groups

test=develop

support heter-xpu-ps

test=develop

* Fix conv1d when data formate is NLC
test=develop

* Fix unitest of conv1d
test=develop

* fix _to_tensor method of Distribution class

* Add unittest

* let dtype be consistent with value in log_prob and probs

* fix format

* fix dtype problem and change unittest

* fix dtype of Numpy class in unittest

* add formula for entropy and kl

* change formula

* fix kl formula format

* fix kl formula format 2

* change gt to np in unittest

* optimize unittest format

* delete dumplicate

* delete dumplicate 2

* extract common function used to convert dtype value

* fix conv output_size has no default value bug
* fix summary bug