wx3000 edited a comment on issue #4102: URL: https://github.com/apache/tvm/issues/4102#issuecomment-784456088
@srkreddy1238 : I agree we should avoid duplication in ops implementation. This will speed things up for TF2.x model support and keep the code clean. Great idea! I think we can agree to have a series of small PR in this order: 1. new entry point from_tensorflow() for v2 style control flow handling, in a new file tensorflow2.py 2. common ops moving from tensorflow.py to tensorflow_ops.py. examples are: _concatV2, _transpose, etc. ops in this file are tested for both versions of the parser. 3. new ops such as TensorList* will be in tensorflow2_ops.py first. Later we may move it to tensorflow_ops.py if there is a need for this op in the v1 parser. 4. common parser entry point for v1/v2 switch based on inspection of the graph: if ops are supported only by v2 parser then choose v2 parser, otherwise use v1 parser. 5. convenient entry point which takes saved_model format or concrete_function, in addition to GraphDef Myself and a few others I know will soon submit item 1-3 from the above list. Rest I don't know who will be working on at this point but hopefully someone will pick it up. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
