Fix keras Conv2D BiasAdd fuse

[hwangdeyu]

fixes #1741.

Signed-off-by: hwangdeyu dejack953@outlook.com

[lgtm-com]

This pull request introduces 1 alert when merging 98f72d4 into 72d6460 - view on LGTM.com

new alerts:

• 1 for Unused local variable

[hwangdeyu]

Hi @guschmue,
after adding the tf_optimize() in `_from_keras_tf1', the optimize() may make it model change. So all the keras test models using model predict are failed.
Do you have any idea to fix this cases?

[guschmue]

tf_optimize() sure will change the graph but should not change numerical results. We call it for every path (just not the one in this issue) - must be some problem how we call it in this case.

[hwangdeyu]

tf_optimize() sure will change the graph but should not change numerical results. We call it for every path (just not the one in this issue) - must be some problem how we call it in this case.

Yeah, there may have been a misunderstanding in my comment before.

I mean that after adding the optimize change code, some tests are failed in model.predict test_api.py#L68 and throw the error:

tensorflow.python.framework.errors_impl.NotFoundError: PruneForTargets: Some target nodes not found: predict/group_deps 

But test_api.py #L65 L66 can be passed. So the onnx graph is ok and onnxruntime can be run successfully and get the same numerical results with original keras model.
Now I'm tring to figure out how the tf_optimize() change the graph and make model predict failed.