stephenrawls commented on issue #14208: Add support for fast variable-length LSTM URL: https://github.com/apache/incubator-mxnet/pull/14208#issuecomment-467258960 @szha Okay I re-worked the code to take in sequence_length on the GPU and copy to the CPU inside the kernel. I added a unit test to test the functionality and it appears to be working. For some reason I am getting errors when trying to pass an int32 sequence_length array. Not sure why, because I changed infer_shape() to allow it. Can look into it later but if anyone has suggestions please let me know. @ptrendx -- I figured out the weird import error and reverted those changes to optimizer.py in my PR. Also as mentioned above, since I am now doing an extra GPU->CPU copy in the kernel code, I have also reverted the change to imperative_utils.h. @szha -- I think all that is remaining in this PR is to add the `CHECK(!use_seq_len)` for the CPU case, to figure out why int32 sequence_length is not working, and any cleanup / modifications you think are appropriate. Also--I added a unit test for gluon, but should I also add one for the Symbol api code, or is it sufficient to test this functionality in Gluon?
---------------------------------------------------------------- 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