Hi @jasonyu1996. Thanks for reporting this. Ive been looking to implement a trace operator in mxnet per this request #10500. In preparing for implementing trace (which is really just summing the diagonal of the matix) I also noticed the limited implementation of the diag operator.
Given that many MXNet users have data in the form of WxHxC (width, height, channel), and then add a 4th dimension for number/batch, what are your thoughts on general N-dimensionality approach for this operator? Is it necessary to support the general case? And how about implementation of the diag operator, as you mentioned there are already existing and high performance implementation for each sub-computation required. Do you think there is opportunity for further performance improvement (memory, time, etc.) by fusing these together? Or do you think it would be best to just implement diag calling these sub-computations separately (inside the diag operator)? Let me know if you have thoughts on this. I would be interested in working with you to implement this as well. Heres the original diag issue: #9253 [ Full content available at: https://github.com/apache/incubator-mxnet/issues/12327 ] This message was relayed via gitbox.apache.org for [email protected]
