Onnx optimizer

Onnx optimizer is a C++ library with a python frontend to perform backend-agnostic optimizations of onnx graphs,.

To contribute an optimization pass we generally need to do the following:

1. Add a new file here

   • subclass the Pass class (or one of its subclasses) as defined here

   • PredicateBasedPass is a very common type of Pass to implement. Once this is subclassed, we need to implement patternMatchPredicate and runTransform

2. Register the new Pass here

3. Add a test here

Some of my contributions to this library are:

• idempotent Sign op optimization: Since "Sign" op is idempotent (applying it more than one time has the same effect as applying it one time), we can add it to the EliminateConsecutiveIdempotentOps pass

• where-not optimization: where(not(b), x, y) is equivalent to where(b, y, x). In such cases we can remove the not op.