* bump up 1.x branch to 1.7.0

* bump version for clojure

* Using unrestricted

* Drop publish step

* Enable restricted nodes

* Reverted website_full, added website_nightly

* Reduced node labels to utility and linux_cpu

…che#18044)

Backport of apache#18018

…18038)

* Support projection feature for LSTM on CPU (Only Inference) (apache#17702)

* Support projection feature for LSTM on CPU

* test solution for -Werror=maybe-uninitialized

* Check device type when create state

* Document the projection feature of LSTM for RNN operator

* Minor fix

* Re-run CI

* Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1 (apache#17872)

* Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1

* Use nd.copy() to initialize parameters of new operator

* Add check for output states

* Initialize i2h/h2h_weights with zeros for rnn_relu/tanh, and reduce size

* Split fused rnn layer test into tests of individual mode

* Skip lstm and gru tests on CPU context without DNNL

* optimize for backward batchnorm

* using memcpy instead of 'for' loop

* rm unnecessary pointer cast and add const for some variable

* trigger CI

… (apache#18033)

* Update 3rdparty/mkldnn remote URL and pin to v1.3 (apache#17972)

* update onednn remote url

* checkout onednn v1.3 release

* fix format test

* make test
Conflicts:
	.gitmodules
	3rdparty/mkldnn
	tests/cpp/operator/mkldnn_test.cc

* build flag

* upgrade cmake

* Temporal solution for fp16 accumulation in Bert gemms

* Resolve alpha/beta type issue

* add documentation for env variable MXNET_FC_TRUE_FP16

* Improve description of env variable

* Add unitest checking environment variable

* keep pseudo-fp16 if architecture does not support Float16Compute

* Fix cpplint

…e#17762) apache#18063 (apache#18069)

* Dynamic subgraph compile support (apache#17623)

This PR adds support for passing the NDArrays from the existing optimize_for API down to the reviewSubgraph function in an external library. It also adds a new API for HybridBlock called optimize_for that can partition the model without running a forward pass.

Feature changes

    Adds new API to HybridBlock optimize_for that partitions the model but does not call the cachedOp
    Modifies the subgraph library example to optionally require args to be provided
    Adds annotation on subgraph inputs for the name of the original param so that inputs can be mapped and passes annotations to input nodes of subgraphs
    Adds support for tensors in MKLDNN format, calls Reorder2Default

New tests

    Adds a new test to partition operators that directly consume params
    add a new model to test where ops to be partitioned have args/params

Bug Fixes

    fixes bug in passing ids vector by value instead of by reference
    fixes bug in passing copies of attributes instead of by reference
    fixes bug where _cached_graph was not updated after partitioning
    fixes memory leak where user-specified attributes on subgraph ops were not freed if subgraph was rejected
    fixes problem incorrectly indexing into shape/dtype maps when annotating the graph

Docs

    Updates the README doc with the latest changes described above

* Adding sparse support to MXTensor for custom operators (apache#17569)

* Added enum for sparse storage

* Add structure for Dense and Sparse

* redesign the data structure for MXSparse

* pull out aux data from sparse NDArray

* Added more sparse arguments to API interface

* Passed sparse from c_api to lib_api.h and set in MXTensor

* Fix indent

* fix segfault

* Fix NDArray to MXTensor errors

* Add a sample of sparse(CSR) transpose

* Make CSR transpose temporarily work by hardcoding

* Fixed sparse output size(Refined)

* Add tests for symbolic and stateful ops

* Added a sample for row sparse transpose

* Added real row sparse transpose

* Fix output size issue by adding lambda for CheckAndAlloc()

* Fix mixed storage formats error

* Added infer storage type function

* resolve comments

* Set inferSType as optional function

* Resolve comments

* Add error messages

* Resolve comments

* verify transpose ops results

* fix sanity check

* update MX_LIBRARY_VERSION to 5

* Custom Operator Random Number Generator Support (apache#17762)

Add random number generator support for custom operator libraries.

Design: We pass from MXNet the initialized and seeded states, located on CPU and GPU, to custom library. So user could use those seeds to generate deterministic values from a given seed passed to MXNet. Basically this workflow:

mx.random.seed(128)
r1 = mx.nd.some_custom_random_op(data)
mx.random.seed(128)
r2 = mx.nd.some_custom_random_op(data)
assert (r1 == r2)

This PR does not let custom library generate exactly the same sequence of random numbers comparing to MXNet

This is a continuation of the custom operator project apache#15921 and apache#17270

Co-authored-by: guanxinq <[email protected]>
Co-authored-by: Ziyi Mu <[email protected]>

…oling backward" (apache#18067)

* [MKLDNN] support using any format in pooling backward (apache#17900)

* use any format in pooling backward

* use data_type()

* fix backport

…iction (apache#17994) (apache#18085)

(cherry picked from commit afae030)

…duce overhead (apache#17707) (apache#18039)

* Add LeakyReLU:Gelu (fwd and bwd) to fused ops

* Add test LeakyReLU:gelu

* cpplint

* fix lint

* fix bug SQRT_2 using constant memory

* add comments

)

* Fix ElemwiseSum for more than 4 inputs (apache#17995)

* Fix ElemwiseSum for more than 4 inputs

* Added test

* Fix for handling negative indices in the fusion of slice (apache#17937)

* Fix for handling of negative axis, begin and end in fusion of slice ops

* Added test

…18064)

* [MKLDNN] apply MKLDNNRun to quantized_act/transpose (apache#17689)

* apply MKLDNNRun to quantized_act/transpose ops

* run CI

* [MKL-DNN] Integrate Conv3d and Pool3d/1d (apache#17884)

* Integrate MKl-DNN conv3d and pool3d/1d

* fix UT & address comments

* clean code

* rebase against latest master

* fix conflicts

* fix CI

* rebase

apache#18095)

* Vectorized loads for binary elemwise kernel

* More generalization

* Add backwardusenone

* Remove the unused _backward_add op

* Add vectorized backwardusein

* Extending vectorization to more binary ops, binary ops with scalar and
unary ops

* Handling ElementwiseSum

* Get rid of half2 in mshadow

* Remove backward_elemwiseaddex

* Revert "Remove the unused _backward_add op"

This reverts commit f86da86.

* Revert "Remove backward_elemwiseaddex"

This reverts commit 7729114.

* Add back the backward_add since C++ test relies on it

* Test bcast implementations

* First version of vecotrized bcast

* Adding single side vectorized bcast kernel

* Removing debug prints

* Actually run the single side kernel

* Move the default implementation of bcast to the vectorized one

* Limit the new implementation to GPU only

* Enabling vectorization when broadcast does not actually do broadcast

* Cleaning

* Cleaning part 2

* Fix for numpy ops using stuff from broadcast

* Fix

* Fix lint

* Try to debug pinv numpy test

* Fix

* Fix the vectorized broadcast implementation for misaligned input
pointers

* Added tests

* Added docs to cuda_vectorization.cuh

* Another fix for broadcast and fix INT64 compilation

* Optimize for aligned=true

* 1 more addition to test

* Reverting the change to Numpy op test

* Trying mcmodel=medium to fix the failure in CMake static build

* Revert "Trying mcmodel=medium to fix the failure in CMake static build"

This reverts commit 1af684c.

* Limiting the PR to just elementwise ops

* add debug prints to debug error in CI

* add debug prints to debug error in CI

* remove prints

* initial commit

* enabled calling create for selector

* connected selector to call external class

* added code to remove temp graph attrs

* fixed build issues

* changed shape inference to use different attr names

* fixed selector class

* cleaned up APIs

* fixed sanity

* updated build for extensions

* sanity fix

* refactored MXLoadLib into separate functions

* undo rebase

* finished merge

* enabled verbose in library loading

* fixed example

* added passing args/aux down to graph pass

* added creating new args/aux for graph passes

* fixed return args/aux

* fixed sanity

* whitespace

* fixed lint

* updated perl API, README, added pass_lib to cmake build flow

* fixed mistake with relu example lib

* fixed perl syntax

* addressed comments

* addressed more comments

* fixed compile issues

Co-authored-by: Ubuntu <[email protected]>
Co-authored-by: Ubuntu <[email protected]>

Co-authored-by: Ubuntu <[email protected]>
Co-authored-by: Ubuntu <[email protected]>

… locale that uses commas as the decimal separator (apache#17177)

* Add a test for floating point parsing locale invariance

* Use locale-invariant dmlc:stod/stof instead of std:stod/stof

* Change the new operator tutorial to use dmlc:stod instead of std::stod

* Rename locale invariance test

* Skip test_scalarop_locale_invariance if the locales aren't available

* Fix linter errors due to incorrect include order

* Add Apache License for mshadow

* update cpp-package license

* udpate license for mx-theme in top-level LICENSE

* Enable RAT License check for mshadow, and keep the rest of 3rdparty unchanged.

* add license header

…ache#18141)

* For mxnet-validation pipeline, require sanity build to complete successfully before running other build pipelines. (apache#17999)

* Refactor staggered builds - create new full build pipeline that runs sanity check first, then starts all other builds.

* Move list of build jobs to top of file for clarity. Preserve whole job path in case we use nested folders in the future.

Co-authored-by: Joe Evans <[email protected]>

* If sanity build is not found, wait until Jenkins recognizes it. (apache#18119)

* If sanity build is not found, wait until Jenkins recognizes it.

* Also add a timeout of 30m for sanity build to run and complete, so we don't get stuck in a loop.

Co-authored-by: Joe Evans <[email protected]>

Co-authored-by: Joe Evans <[email protected]>

…pache#17772) (apache#18075)

Co-authored-by: Ubuntu <[email protected]>

Co-authored-by: Ubuntu <[email protected]>

* Fix pylint astroid sanity issue

Cherrypick apache@18e2014 from apache#18220

* Update edge toolchain

* Support platforms without rand_r

* Fix the URL to the IUS repository

* compiler warnings

* Use a pre-c++17 way of distinguishing between device types

* Greatly simplify qemu setup

* Request the C++ standard library and extensions

* Upgrade dmlc-core to resolve build errors

* Remove leftovers from C++17 dev type check

* Fix CPU-only RRNOp Forward

* Change the ARM8 build to work like the ARM7 build

* Revert "Fix CPU-only RRNOp Forward"

This reverts commit 0a921a4.

* Hack around the lack of constexpr if

* Adjust the list of files to be packed in ARM jobs

Co-authored-by: Leonard Lausen <[email protected]>

Fix leak of ndarray objects in the frontend due to reference cycle.

Backport of 3e676fc

* fixed overwrite of args/aux variables

* fixed spacing

* Merged apache#18177

* updated python RPM URL

…apache#18309)

* Revert "Fix and optimize handling of vectorized memory accesses (apache#17767)"

This reverts commit 5542d03.

* add license to reverted file

) (apache#18432)

* remove OS from s3 library path

* fix bash script run commands

* Revert "remove OS from s3 library path"

This reverts commit 2665113.

* hardcode s3 path for upload/download of binaries

Co-authored-by: Ubuntu <[email protected]>

Co-authored-by: Manu Seth <[email protected]>
Co-authored-by: Ubuntu <[email protected]>

* add bnrelu bf16 into amp list

* [MKL-DNN] BatchNormRelu Fusion (apache#17679)

* support bnrelu fusion

* add param to gluon

* move to contrib

* fix gluon interface

* reuse bn param

* fix

* fix forward

* cache flags

* fix lint

* fix ut

* trigger

* trigger

* trigger

* inherit from BN base

* fix lint

Co-authored-by: no <[email protected]>

* Remove chinese period which leads to utf-8 encoding problem (apache#18223)

* add bnrelu amp test

* trigger ci

Co-authored-by: no <[email protected]>
Co-authored-by: damNull <[email protected]>

…upport (apache#18320) (apache#18469)

* Improve log_softmax performance by OneDNN library

* Adapt tests for MKLDNN log_softmax

* Fix lint errors

* Fix indent and comments

Update basic_layers.py

fix

fix

Update basic_layers.py

fix bug

Co-authored-by: Xingjian Shi <[email protected]>

…17632) (apache#18317)

* [v1.x] [Large Tensor] Backport of Fixed RNN op (apache#17632)

* Changed relevant function args to index_t

* Added nightly test for RNN

* Added fix for LSTM, GRU, RNN-ReLU, RNN-tanh

* Using const instead of literals

* Added nightly test for RNN ReLU & tanh, LSTM, GRU

* Type assertion to force evaluation of output NDArray

* Incorporated latest round of comments

* [v1.x] Backport of Fix LSTM and GRU layers gradient calculations (apache#18203)

* Fix input gradient calculation for bidirectional LSTM

For bidiractional LSTM with number of layers > 2 input gradient calculation was incorrect.
Reason of wrong calculations was overwriting y derivative (dy) tensor by
calculated x derivative (dx) tensor before right2left layer could use dy for own
gradient calculations.
Propsed fix uses additional space to avoid overwriting.

* Fix gradient calculation for GRU

For GRU with number of layers > 2 i2h_weight gradient for
layers in the middle (all except last and first) was incorrect.
Wrong caluculations were caused by assigning output pointer to
input instead of calculating new input pointer.

* Enable tests for GRU and LSTM gradients

* Fix comments

* Change loop iteration deduction

* Add more test cases for fused rnn layers

Co-authored-by: Connor Goggins <[email protected]>

…pache#18503)

…eq is `add` (apache#18518)

* Fix batch norm when grad_req is

* fix

* remove softmax test

* fix

* Fix CD (apache#17776)

* Fix cd/mxnet_lib/dynamic/Jenkins_pipeline.groovy

Fixes a regression in CD introduced by
apache#17645

* Fix whitespace

* Add NATIVE_ADDITIONAL.md

Fixes a regression in CD introduced by
apache#16899

* Update other $TYPE_ADDITIONAL.MD

* Fix cd/python/docker

Fixes regression introduced by
apache#15990

* [CD] update pypi description, setup.py (apache#17681)

* update pypi description, setup.py, use manylinux2010, use unified dist link for nightly

* Use manylinux2014

Co-authored-by: Leonard Lausen <[email protected]>

* reverting .so path as per MAKE flow

Co-authored-by: Leonard Lausen <[email protected]>
Co-authored-by: Sheng Zha <[email protected]>
Co-authored-by: Leonard Lausen <[email protected]>

* add config Makefile for jetson

* modify jetson install guide

* Increase staggered build timeout to 180 min, since sanity build has 180 min timeout.

* Decrease timeout so everyone is happy.

Co-authored-by: Joe Evans <[email protected]>

…apache#18634)

Co-authored-by: Leonard Lausen <[email protected]>

…) (apache#18654)

* Add test for BatchNorm running variables synchronization

* Fix BatchNorm backward synchronization

It fixes issue apache#18610

…le input shapes (apache#18705)

* Migrate from private to public jetson toolchain files (apache#18677)

* Set CMAKE_CUDA_COMPILER in aarch64-linux-gnu-toolchain.cmake (apache#18713)

CMAKE_CUDA_HOST_COMPILER will be reset if CMAKE_CUDA_COMPILER is not set as of cmake 3.17.3

See https://gitlab.kitware.com/cmake/cmake/-/issues/20826

Co-authored-by: Leonard Lausen <[email protected]>

…18768) (apache#18778)

…#18752)

* Fix linalg_potri operator for large tensor.

* Update other variables to support large tensors.

* Add to contributors.

* Fix whitespace.

* Update ZeroTriangular to support large tensors.

* Add large tensor unit tests for linalg_potrf and linalg_potri.

* Fix crash when accessing already destructed static variables (apache#18768) (apache#18778)

Co-authored-by: Joe Evans <[email protected]>
Co-authored-by: Przemyslaw Tredak <[email protected]>

* add large tensor test for syrk, foward and backward

* change to batch input

* move syrk test into test-linalg

Co-authored-by: Ubuntu <[email protected]>

apache#18744)

* add linalg large matrix tests

* add batch inputs linalg tests

* reducing bsize to 1 to save time

* move matrix generator to utils

* passing mat size as arg

* import util fn

* fix sanity

* add mx

* call backward

* merge fn

* update grad value

* refactor tests

* add mx

* add shape check

Co-authored-by: Ubuntu <[email protected]>

* added forward, backward test for gemm2

* add backward check

* correct gradient assert

* move test inside linalg_ops

* add shape checks

* Improving performance of broadcast_axis on GPU (apache#18168)

* adding separate int32_t kernel for GPU in broadcast_axis/to/like operators

* using structure instead of temp workspace to pass stride and shape

* replacing hardcoded int32_t with generic index_t

* combining CPU and GPU kernels to leverage cached stride calculation and fast access shape data in both

Co-authored-by: Rohit Kumar Srivastava <[email protected]>

* Improve performance of broadcast_axis on CPU (apache#17882)

* adding comments explaining code optimizations

* fixing broadcast_axis kernel to int32

* fixing slice_axis kernel to int32

* combining CPU and GPU implementation method signatures and cleaned up
code

* adding new broadcast_axis to np_matmul

Co-authored-by: Rohit Kumar Srivastava <[email protected]>

Co-authored-by: Rohit Kumar Srivastava <[email protected]>

* add shape check

* add name to contributor.md

Co-authored-by: Ubuntu <[email protected]>

…syevd (apache#18807)

Co-authored-by: Rohit Kumar Srivastava <[email protected]>

…ache#18825)

* add test for linalg.gemm

* fix indents

* initial

* test

* gemm and gemm2

* type fix

* syrk trmm trsm

* gelqf

* move tests from test_large_array.py to test_large_vector.py

* fix white space issue

Co-authored-by: Ubuntu <[email protected]>

… unit tests. (apache#18803)

Co-authored-by: Joe Evans <[email protected]>

…he#18602) (apache#18708)

* [v1.x] Backport of fix npx.softmax for 0-sized inputs (apache#18158)

Co-authored-by: Hao Jin <[email protected]>

* Fix softmax, logsoftmax failed on empty ndarray (apache#18602)

* Fix failing empty array (log_)softmax

* Modify test for npx (log_)softmax

* Fix softmax, logsoftmax backward failed on empty ndarray (apache#18710)

Co-authored-by: Yiyan66 <[email protected]>
Co-authored-by: Hao Jin <[email protected]>
Co-authored-by: Bart Gawrych <[email protected]>

…che#18846)

* fixing spatial export for batchnorm

* retrigger CI

* fixing broken pylint

* retrigger build

* deprecating spatial attribute in exporter so default behavior of spatial=1 is conveyed

Co-authored-by: Vinitra Swamy <[email protected]>

* Cherry-pick apache#18310 apache#18355 (apache#18608)

* cherry-pick: Fix missing MKLDNN headers (apache#18310)

* Include all mkldnn headers in CD builds (apache#18355)

* Fix cmake mkldnn install target. Previously mkldnn headers are installed to CMAKE_INSTALL_INCLUDEDIR instead of CMAKE_INSTALL_INCLUDEDIR/mkldnn

* Fix pypi_package.sh pip/setup.py for mkldnn builds

* Set CMAKE_CUDA_COMPILER in aarch64-linux-gnu-toolchain.cmake (apache#18713)

CMAKE_CUDA_HOST_COMPILER will be reset if CMAKE_CUDA_COMPILER is not set as of cmake 3.17.3

See https://gitlab.kitware.com/cmake/cmake/-/issues/20826

Co-authored-by: Leonard Lausen <[email protected]>

* remove linux-gputoolchain

Co-authored-by: MoisesHer <[email protected]>
Co-authored-by: Leonard Lausen <[email protected]>

apache#18867)

* Update mirror for getting binutils source.

* Remove erroneous wget command and duplicate mkdir command.

Co-authored-by: Joe Evans <[email protected]>

…o v1.7.x (apache#18676) (apache#18890)

* [Improvement] Invoke mkldnn and cudnn BatchNorm when axis != 1 (apache#18504)

* fix batch norm when fix_gamma is True

* support gradient accumulation for batch norm

* mkldnn batchnorm support grad add

* unittest for bn

* fix bn arg

* fix lint

* fix mkldnn

* fix mkldnn bn

* fix grad when fixing gamma

* fix naive gpu bn

* fix lint

* invoke mkldnn and cudnn batchnorm when axis != 1

* backport 18500

* change condition

* fix

* fix

* add mkldnn_off for bn

* remove mkldnn_off

* recover save_000800.json

* cast

* remove  and fix flaky test

Co-authored-by: JackieWu <[email protected]>

Co-authored-by: JackieWu <[email protected]>

…apache#18895)

Signed-off-by: Serge Panev <[email protected]>

* Remove mention of nightly in pypi (apache#18635)

* update bert dev.tsv link

Co-authored-by: Sheng Zha <[email protected]>

Co-authored-by: Carin Meier <[email protected]>
Co-authored-by: Sheng Zha <[email protected]>

Co-authored-by: Tao Lv <[email protected]>

…he#18785)

* Update unix gpu toolchain (apache#18186)

* update nvidiadocker command & remove cuda compat

* replace cu101 with cuda since compat is no longer to be used

* skip flaky tests

* get rid of ubuntu_build_cuda and point ubuntu_cu101 to base gpu instead of cuda compat

* Revert "skip flaky tests"

This reverts commit 1c720fa.

* revert removal of ubuntu_build_cuda

* add linux gpu g4 node to all steps using g3 in unix-gpu pipeline

* remove docker compose files

* add back the caffe test since caffe is deprecated for mx2.0 and not 1.x

* drop nvidia-docker requirement since docker19.0 supports it by default

:q

* remove compat from dockerfile

* Cherry-pick apache#18635 to v1.7.x (apache#18935)

* Remove mention of nightly in pypi (apache#18635)

* update bert dev.tsv link

Co-authored-by: Sheng Zha <[email protected]>

* disable tvm in CI functions that rely on libcuda compat

* tvm off for ubuntu_gpu_cmake build

* drop tvm from all unix-gpu builds

Co-authored-by: Carin Meier <[email protected]>
Co-authored-by: Sheng Zha <[email protected]>

* initial commit

* Support extra inputs for subgraph ops (apache#18779)

Support additional inputs to custom subgraph ops that are not direct dependencies to ops in the subgraph. This will enable various use cases: custom control flow ops, custom ops that maintain a state that should be saved/loaded, etc.

Highlights:

* Added test that uses a graph pass (addInputPass) to add a new custom input to the subgraph op

* Added new optional argument (clear) to hybridize & optimize_for APIs in Gluon Block to enable multiple optimizations

* refactored lib_api.h JSON utilities

* added new Graph data structure utilities to simplify custom graph passes

* refactored custom op registration

* enhanced custom subgraph op to support additional inputs to subgraph op that is not an input to ops in the subgraph

* updated subgraph & graph pass READMEs

* Added error messaging from external library

* changed messages

* changed to pointers and types

* added cast

* updated cast

* fixed signed int

* whitespace

* fixd pass resource

Co-authored-by: Ubuntu <[email protected]>

…t) (apache#18916)

* [1.x] Backporting TensorRT and Gluon changes

Signed-off-by: Serge Panev <[email protected]>

* Remove test from Jenkins

Signed-off-by: Serge Panev <[email protected]>

* Fix test

Signed-off-by: Serge Panev <[email protected]>

…ache#18929 (apache#18964)

Signed-off-by: Serge Panev <[email protected]>

) (apache#18973)

Co-authored-by: Vladimir Cherepanov <[email protected]>

Co-authored-by: Vladimir Cherepanov <[email protected]>

… (apache#19009)

* Fix LeakyRelu behaviour on empty input

* Remove duplicated declarations

* Stop packaging GPL libquadmath.so (apache#19053)

libquadmath.so is GPL and must not be distributed by Apache projects. Users will need to ensure that libquadmath.so is present on their systems if they use binary builds of MXNet. libquadmath.so has not yet undergone any ABI changes, thus all versions of libquadmath.so are ABI compatible and user just needs to install system version of libquadmath.so.

libgfortran.so can be packaged thanks to GCC Runtime Library Exception.

See https://www.apache.org/legal/resolved.html#category-x

* Remove unmaintained pip packages

* Workaround pypa/setuptools#2352

Backported apache#18930

* Support for fp16 in SpGeMM

* adding test for GPU spmm

Co-authored-by: Rohit Kumar Srivastava <[email protected]>

Co-authored-by: Rohit Kumar Srivastava <[email protected]>

* initial commit

* fixed c++17 downgrade

* fixed stringstream

* fixed cast

* changed to use pointers for stringstream since not copyable

* fixed includes

* fixed makefile includes

* skipped lint for malloc/free for passing across C ABI

Co-authored-by: Ubuntu <[email protected]>

This PR cherry-picks commit 5122d32 into the v1.x branch. This is to enable the export of models where dangling layers are optimized out during symbol export. For more information, see here and here.

* empty list cannot be cleared issue fixed.

* Update multiproc_data.py

Co-authored-by: Sheng Zha <[email protected]>

Signed-off-by: Serge Panev <[email protected]>

apache#19017)

* fix pooling_convention warning when convert model to onnx (apache#18529)

* fix  pooling_convention warning

* fix pooling_convention warning

* fix lint

Co-authored-by: JackieWu <[email protected]>

* Prevent uninitialized variable error.

* Initial work to get Dropout to work with onnx 1.7

* Remove trailing whitespace for pylint.

* Fix tensor initialization for Dropout operator input.

* Update Clip operator to support latest ONNX opset versions by moving min/max attributes to inputs.

* Fix whitespace.

* Add support for importing Dropout operator in ONNX opset version >= 12.

* Add support for import ONNX opsets >= 11 to clip operator.

* Add optional opset_version parameter that defaults to latest opset version supported by onnx. Pass this parameter to each graph layer when exporting.

* Add optional parameter to create_model() that allows user to specify which onnx opset version they want to use when exporting, defaults to latest version supported by onnx.

* Use opset_version argument to determine operator format.

* Add a opset_version parameter to from_onnx() so at operator conversion time, we know what opset version to use.

* For Clip and Dropout operators, use opset version from passed proto_obj, which reflects what opset version the onnx model uses.

* Use same tolerances that are in master.

* Change Pad operator to use inputs instead of attributes for newer opset versions. Check opset version instead of ONNX version for Pooling operator.

* Add documentation opset_version parameter.

* Add opset_version parameters to unit tests.

* Add test script for testing inference with onnxruntime on CV models from gluon model zoo.

* Add license and clean up imports.

* Install onnxruntime in docker container for unit tests.

* Add onnxruntime to test dependencies.

* Install onnxruntime into CentOS docker image.

* Disable testing squeezenet models for now.

* Update onnx version.

* Fix typo.

* Use mx.image.imread instead of PIL module.

* ONNX import: use Conv pad attribute for symmetrical padding (apache#18675)

Signed-off-by: Serge Panev <[email protected]>

* Install onnx in CentOS containers when installing python.

* Update import and export of some ONNX ops to support newer opset versions - this gets all ONNX unit tests to pass with onnx 1.7.

* Re-enable squeezenet model testings in onnxruntime.

* Run the onnxruntime inference tests in the ONNX pipeline instead of normal unittests pipelines.

* Add missed return value.

* Refactor code based on review comment.

* Since the onnx tests are only run on ubuntu_cpu images, we don't need to install onnx and onnxruntime in the CentOS containers.

Co-authored-by: Liu, Hao <[email protected]>
Co-authored-by: JackieWu <[email protected]>
Co-authored-by: Joe Evans <[email protected]>
Co-authored-by: Serge Panev <[email protected]>

…9122)

* Wait for async_fun to complete in NaiveEngine::PushAsync

This fixes a race condition in which NaiveEngine::PushAsync was checking if the
the async_fun had completed by the end of NaiveEngine::PushAsync. If async_fun
hadn't completed yet, NaiveEngine::PushAsync would set an internal error string
and deallocate the callback, causing segfault in async_fun once it would attempt
calling the callback.

* Update naive_engine.cc

* initial commit

* incremented version number

Co-authored-by: Ubuntu <[email protected]>