zxybazh opened a new pull request, #14324: URL: https://github.com/apache/tvm/pull/14324
In the case where the begin index of strided slice is out of [-ndim, dim), the strided slice operator will yield some incorrect shape inference. This PR corrected this issue by canonicalizing the begin and end index for strided slice and calculate the symbolic shape. Added some new unit tests for out of range begin locations and changed some symbolic shape tests. I have 2 questions that we can follow up in the design: 1. Is it expected that in the case of symbolic shape we should have symbolic shape inferenced after strided slice? See unit test change for details. 2. We do have the similar function / logic in topi strided slice implementation. Should we unify them some how? Is it expected that we use some mechanism to store the shape inference logic for both relax / topi operators? Would be great to see other contributors opinions on this. CC: @MasterJH5574 @sunggg @tqchen -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
