roywei opened a new issue #14833: [Numpy] Infer shape partial failed for 
unknown shapes
URL: https://github.com/apache/incubator-mxnet/issues/14833
 
 
   I came from a user reported issue here: 
https://github.com/apache/incubator-mxnet/issues/14751#issuecomment-486866935
   
   In https://github.com/apache/incubator-mxnet/pull/14661, unknown shape in 
mxnet was represented using `0`, and changed to `-1`, this brought some 
unexpected behavior on `infer_shape_partial` of some of the operators in MXNet. 
 
   As noted in the [PR](https://github.com/apache/incubator-mxnet/pull/14661) 
discussion, when many operators were implemented for the first time, some 
strong assumption may have been made because zero-dim and zero-size tensors are 
not introduced.
   
   Also, many of these problems are not revealed in MXNet because unknown dim 
place holders (0 or -1) are not used often. For now, this change does not bring 
any other compatibility issue except in the 
[keras-mxnet](https://github.com/awslabs/keras-apache-mxnet) project. 
Keras-MXNet relies a lot on unknown shapes and infer_shape_partial internally.
   
   As I'm trying to fix the problem, I m using this issue to record the 
operators in MXNet where the infer shape logic may be wrong. 
   
   Following are the operators I found now, more to come:
   - [ ] dot
   
   - [ ] transpose
   
   cc @reminisce @junrushao1994 @sandeep-krishnamurthy 

----------------------------------------------------------------
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]


With regards,
Apache Git Services

Reply via email to