[GitHub] HuichuanLiu commented on issue #11149: Unreasonable performance of resnext models provided in model_zoo, evaluated by score.py

[GitBox <gi...@apache.org>]

HuichuanLiu commented on issue #11149: Unreasonable performance of resnext models provided in model_zoo, evaluated by score.py
URL: https://github.com/apache/incubator-mxnet/issues/11149#issuecomment-394693892
 
 
   I take a look at the symbol.json file from [model_zoo](http://data.dmlc.ml/models/imagenet/resnext/101-layers/resnext-101-64x4d-symbol.json), it looks very different from the original paper.
   In my understanding, the resNeXt feeds every path with the same input then the elemwise_add is conducted cross all 32 paths/units and after that the residuals are finally involved. As the figure shows
   ![image](https://user-images.githubusercontent.com/10973111/40976040-73f92a52-68ff-11e8-9c2d-1116e9adcf8b.png)
   
   But according to the symbol.json , the Nth units receive outputs from the N-1 th unit, and the residual element-wise add operation is conducted after each unit, one by one. Is it some trick in implementation?
   

----------------------------------------------------------------
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:
users@infra.apache.org


With regards,
Apache Git Services