Xingjian Shi created MXNET-31:
---------------------------------
Summary: Support variable sequence length in gluon.RecurrentCell
Key: MXNET-31
URL: https://issues.apache.org/jira/browse/MXNET-31
Project: Apache MXNet
Issue Type: New Feature
Reporter: Xingjian Shi
Assignee: Chris Olivier
When the input sequences have different lengths, the common approach is to pad
them to the same length and feed the padded data into the recurrent neural
network. To deal with this scenario, this PR adds a new {{valid_length}} option
in {{unroll}}. {{valid_length}} refers to the real length of the sequences
before padding. When the {{valid_length}} is given, the last valid state will
be returned and the padded portion in the output will be masked to be zero.
This feature is essential for implementing a NMT model.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
