ONNX export: Add Flatten before Gemm

[vandanavk]

Description

ONNX's spec specifies that Gemm input data should be 2D, but this wasn't the case for VGG and ResNet models trained in MXNet and exported to ONNX. This PR adds a flatten node before Gemm to make the input data 2D.

Related issues: onnx/models#90, onnx/onnx#1101, onnx/models#91 (comment)

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Add a flatten node before Gemm

Comments

• Testing done:

1. Exported all VGG models (https://github.com/onnx/models/tree/master/models/image_classification/vgg) and ResNetv1 models (https://github.com/onnx/models/tree/master/models/image_classification/resnet) and executed the validation notebook.
2. A test in mxnet_export_test for fully connected

[anirudhacharya]

how are we testing this change?

[vandanavk]

@anirudhacharya the test is WIP (added a note in the PR description). Adding a test for fully connected in mxnet_export_test.

[vandanavk]

@mxnet-label-bot add [ONNX, pr-work-in-progress]

[vandanavk]

@mxnet-label-bot update [ONNX, pr-awaiting-review]

[nswamy]

@Roshrini, @anirudhacharya for review?

[Roshrini]

Tested this fix on Arcface model in model zoo, LGTM

[nswamy]

@vandanavk
Please review the CI failure

  _, out_shapes, _ = sym.infer_shape(**inputs)

Traceback (most recent call last):

  File "tests/python-pytest/onnx/export/mxnet_export_test.py", line 486, in <module>

    test_models("bvlc_googlenet", (1, 3, 224, 224), (1, 1000))

  File "tests/python-pytest/onnx/export/mxnet_export_test.py", line 142, in test_models

    converted_model_path = onnx_mxnet.export_model(sym, params, [input_shape], np.float32, onnx_file)

  File "/work/mxnet/python/mxnet/contrib/onnx/mx2onnx/export_model.py", line 87, in export_model

    verbose=verbose)

  File "/work/mxnet/python/mxnet/contrib/onnx/mx2onnx/export_onnx.py", line 211, in create_onnx_graph_proto

    graph_outputs = MXNetGraph.get_outputs(sym, params, in_shape, output_label)

  File "/work/mxnet/python/mxnet/contrib/onnx/mx2onnx/export_onnx.py", line 152, in get_outputs

    assert len(out_shapes) == len(out_names)

[codecov-io]

Codecov Report

Merging #13356 into master will increase coverage by 0.53%.
The diff coverage is 0%.

@@            Coverage Diff             @@
##           master   #13356      +/-   ##
==========================================
+ Coverage   82.73%   83.26%   +0.53%     
==========================================
  Files         721      746      +25     
  Lines       79494    83497    +4003     
  Branches     3193     3193              
==========================================
+ Hits        65770    69526    +3756     
- Misses      12864    13113     +249     
+ Partials      860      858       -2

Impacted Files	Coverage Δ
...hon/mxnet/contrib/onnx/mx2onnx/_op_translations.py	18.55% <0%> (-0.14%)	⬇️
src/operator/nn/depthwise_convolution_tf.cuh	83.52% <0%> (-2.36%)	⬇️
src/operator/contrib/quadratic_op-inl.h	96.84% <0%> (-2.11%)	⬇️
src/operator/softmax_output-inl.h	59.19% <0%> (-0.9%)	⬇️
python/mxnet/visualization.py	44.94% <0%> (ø)	⬆️
tests/cpp/kvstore/gpu_topology_test.cc	97.28% <0%> (ø)
tests/cpp/operator/batchnorm_test.cc	87.07% <0%> (ø)
tests/cpp/operator/runner/core_op_runner_test.cc	95.86% <0%> (ø)
tests/cpp/operator/fully_conn_perf.cc	94.73% <0%> (ø)
tests/cpp/include/test_core_op.h	95.23% <0%> (ø)
... and 23 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ba02bf2...739f181. Read the comment docs.

[vandanavk]

@zhreshold for review

[Roshrini]

@zhreshold Can you please review/merge this PR?

[vandanavk]

@mxnet-label-bot update [ONNX, pr-awaiting-merge]