Generate tests for MeanSquareDistance and SoftmaxCrossEntropyLoss #2623
Conversation
|
Maybe a check could be added to CI, similar to how (I believe) it checks docs are up to date |
|
exactly, I like the idea on adding a check on the doc update, same as proto and operator docs. @TMVector are you interested in it? :) |
This should partly fix a bug where names are mapped differently for node inputs and outputs.
|
Unfortunately generating these tests seems to expose issues with the definition of MeanSquareDistance and SoftmaxCrossEntropyLoss, plus with function expansion. I'm looking at fixes now. |
This hopefully fixes a bug related to intermediate values inside function bodies being erroneously dropped if they were also outputs.
| @@ -22,22 +22,28 @@ | |||
| from onnx.onnx_operators_pb import FunctionProto | |||
|
|
|||
|
|
|||
| # FIXME(TMVector): Any reason we can't get rid of this and use the C++ helper directly? | |||
There was a problem hiding this comment.
TODO/question -- can we remove the Python function expansion helper and use the one from C++?
There was a problem hiding this comment.
The changes look good to me, thanks! I guess we can import the C++ version into Python (provided there is no subtle difference between them due to bugs, etc., I suppose). I guess that could be a different PR, given that we are close to the cutoff date for the next ONNX release.
50474e8 to
6443e3d
Compare
6443e3d to
1808901
Compare
…nx#2623) * Generate tests for MeanSquareDistance and SoftmaxCrossEntropyLoss * Fix logic error in MSD/SCE function builders * Regenerate MeanSquareDistance and SoftmaxCrossEntropyLoss backend tests * Add keepdims=0 to Reduce operators in function body of MSD * Add shape inference test for reducing all dims with keepdims=0 * Unify input/output name mapping in function expansion This should partly fix a bug where names are mapped differently for node inputs and outputs. * Use fresh name for missing outputs of functions This hopefully fixes a bug related to intermediate values inside function bodies being erroneously dropped if they were also outputs. * Regenerate MeanSquareDistance and SoftmaxCrossEntropy backend tests * Fix SCE to use 'reduction' attribute from function for NLL * Regenerate backend tests for SoftmaxCrossEntropyLoss Co-authored-by: Prasanth Pulavarthi <prasantp@microsoft.com>
I noticed some backend tests on master hadn't been generated. Looks like the function-expanded ones for #2570 and #2573, so should probably go in for 1.7.