ptrendx commented on issue #9774: does not respect dtype 
   Engine does not seem to differentiate between first layer and subsequent 
layers on that it considers data going into first layer as being modified by 
the backward pass of the network (even though it does not actually happen). 
This means that copying of the next batch has to wait to the end of backward, 
which effectively exposes the copy. Having this double-buffering scheme of 
either cast fp32->fp16 or just identity fp32->fp32 makes sure that the ndarray 
used to copy the next batch to the gpu is returned from engine before backward 
pass ends, which enables copy to happen while backward computation takes place.

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:

With regards,
Apache Git Services

Reply via email to