taliesinb commented on issue #8949: New layer: split_like.
URL: https://github.com/apache/incubator-mxnet/pull/8949#issuecomment-350414149
 
 
   @piiswrong what name makes more sense to you than `split_layer`? perhaps 
`unflatten_layer`, because it is doing the opposite of flattening the first two 
dimensions? I am happy with any name, what is important to me is that it 
prevents the need for multiple unrollings, which is incredibly expensive. 
basically, i don't see how a high level framework that wants to efficiently 
implement variable-length sequences can live without this operation.
   
   i have an alternative to creating a new layer, which is adding more numeric 
codes to the Reshape layer, basically -10, -11, -12, etc which means "take the 
dims from an optional second reference shape input". unfortunately having an 
optional second input in Reshape seemed to be impossible due to technical 
issues i didn't fully understand to do with gradients having the wrong shape. 
also, it seemed like a complicated feature that was only going to be used for 
this splitting purpose. i now regret those previous codes we added to Reshape 
layer, because i have only ever used them for splitting and merging the first 
two dimensions, i think it would be much more understandable to have 
`flatten_dims` and `unflatten_dims` that are only for this purpose. 

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org


With regards,
Apache Git Services

Reply via email to