[GitHub] [tvm] chunit-quic commented on pull request #13728: [Relay][Frontend] Span Filling TensorFlow 1

["chunit-quic (via GitHub)" <gi...@apache.org>]

chunit-quic commented on PR #13728:
URL: https://github.com/apache/tvm/pull/13728#issuecomment-1707891633

   Hi @xiaolong18 
   
   Thank you for your information. Currently we are doing some urgent projects, and perhaps we can only get back to here at the end of October.
   
   But we might give you some hints about this problem. we saw some duplicated ops are generated in the pytorch frontend before. 
   It is caused by a frontend operator which is converted to multiple exprs, and one of these expressions (let's say `p` expr) are comsumed by multiple users(`u1`, `u2`). When span filler recursively traverse back, it construct a set of span-included exprs from `u1` expr until no expr or encountering a span-filled expr. After traversed `u1`, a span-filled `p` is constructed (`sf-p`). However, when spanfiller traverses `u2`, it still found its parent is `p`, rather than `sf-p`. Then a set of duplicated exprs are generated unfortunately.
   
   We handle this kind of problems via invoke set_span directy inside this kind of op conversions, like the `NMS` operator in pyotorch. Perhaps it could be a reference to you.
   
   Hope it could give you clues to do some more investigation. Please feel free to ping us, yet it might take some time for us to get back to you. :D
   
   Thank,
   Joey
   


-- 
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.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org