* dump lowered ir when debug log

* avoid calling tvm.lower() twice when not debug

f9ed337552a26cc55c7c0fb22cd54839ee13f19c added a call to std::log
without including cmath

At present, the build system assumes that the Unix archiv utility 'ar'
will always be present in PATH, but this might not be true if a separate
toolchain is being used. This patch makes the build system a bit more
cross-compile capable.

* Add creating gradient symbol

* Fix lint

* Address comments

* Fix typo

* Address comment

* Update docs

* Add backward operator to enable backward graph

* Fix testing

* Refactor top level1 test code

* Fix format

* Test

* Added zeros ones op

* Register fill_like operator

* Fix unit test

* Add compile library tutorial

* Clean output

* Refactor with sphinx gallery

* Refactor

* Change title and other minor fixes

* keras frontend tutorial

* fix

* vgg16 success

* remove six.PY2, use sys.version_info;
convert_activation() accepts activation type name(str, e.g. 'relu') as input;

* add convert_merge

* fix convert_batchnorm;
improve tests

* fix lint

* add numpy-style pad operator

* deal with asymmetry padding

* resnet50 success

* fix pool_convert; xception passes test

* update tvm

* fix bias error; all tests pass

* use > >, not >>

* [SYMBOL] Add __iter__ and GetChildren for symbol

* [SYMBOL] Fix lint

* Symbol __getitem__ using list_outputs() is too expensive, when it only cares about the output count in most cases

* Add cython cmake

* GetNumOutputs() and __len__ changes per PR comments

* set commit for tvm

* fix mxnet multi outputs

* add test case for multi outputs

* fix

* fix

* fix

* fix index

* use json hack

* fix test cases

* fix test cases

* fix test cases

* fix test cases

* fix test cases

* fix test cases

https://apple.github.io/coremltools/generated/coremltools.models.neural_network.html?highlight=permute#coremltools.models.neural_network.NeuralNetworkBuilder.add_permute

* [CMPL] Add Support for Other Data Types

* [CMPL] Add test

* [CMPL] Fix

Variable declaration correction.

* Make sure to install the nnvm_compiler library.

* Add conda packge.

* Update deploy.md

Add an example in c++.

* Update deploy.md

Add an example in c++.

* Update deploy.md

Add an example in c++.

* Update deploy.md

Add an example in c++.

* Update deploy.md

Update dlpack device and type enumerations.

* update tvm

* [TOP] fix weight layout in conv2d_transpose

* Update deploy.md

Add an example in c++.

* Update deploy.md

Add an example in c++.

* Update deploy.md

Add an example in c++.

* Update deploy.md

Add an example in c++.

* [TOP] add conv2d_transpose

* update tvm

* fix pylint

* update

* fix

* use generated onnx model

* fix tests

* fix lint

* remove log filter

* add vgg

* fix tests

* update tests

* fix download

* fix ci

* fix tutorial url

* clean cache

* graph.input can be either ValueInfoProto or string

* pylint

* Change the way to import ONNX model.

* Decide ONNX version and change the processing

* Check whether onnx has the attribute '__version__'

* Correct the use of `concatenate` operator

* Optimize