[Relay][Frontend][ONNX] New Operators and Opsets to Support BERT (#4197)
authorJosh Fromm <jwfromm@uw.edu>
Wed, 30 Oct 2019 18:24:47 +0000 (11:24 -0700)
committerJared Roesch <roeschinc@gmail.com>
Wed, 30 Oct 2019 18:24:47 +0000 (11:24 -0700)
commit156aa5900ab04d9176cd333bb7d1ce10dce19faa
tree309f56e61273bf8529874744e40f2d0bfff4c20a
parent71f39be5a30eed93fa5f03484d922ec4a34c722e
[Relay][Frontend][ONNX] New Operators and Opsets to Support BERT (#4197)

* Added slice v10

* Added constantofshape operation and small refactor.

* Finished one_hot implementation.

* Reshape working across all bert layers.

* Fixed constantofshape and removed code duplication.

* onnx model fully ingested.

* Working on improving onnx tests.

* Changed onnx testing to use onnxruntime instead of caffe2, also formatted.

* Add arbitrary output nodes to onnx frontend.

* Added v6 tiling for bert squad 8 support.

* Small syntax fixes

* Reduced code duplication in split opset versions.

* Added batch matmul test

* Added unstack split testing.

* Adde onehot test, needs a little cleanup probably.

* Replaced deprecated constant fill with constantofshape and updated tests accordingly.

* Added tests for new opset version of slice and tile.

* lint clean up

* Lint fixes

* Changed onnx dependency

* Went back to caffe2 runtime for CI integration.

* Rebase and small typo/syntax changes.

* Added hard casting of onehot attributes to int.
python/tvm/relay/frontend/common.py
python/tvm/relay/frontend/onnx.py
python/tvm/relay/frontend/tensorflow.py
tests/python/frontend/onnx/test_forward.py