zhangzhaoqi created SINGA-506:
---------------------------------
Summary: add autograd operators for NLP models
Key: SINGA-506
URL: https://issues.apache.org/jira/browse/SINGA-506
Project: Singa
Issue Type: New Feature
Reporter: zhangzhaoqi
*We are going to support these three NLP models, called, Bidirectional
Attention Flow, BERT-Squad and GPT-2.*
*Totally, there are still 21 operators that we need to add as following,*
ArgMax
Cast
CategoryMapper
Ceil
Compress
ConstantOfShape
Dropout
Gather
Hardmax
Identity
NonZero
ReduceMax
ReduceMean
ReduceSum
Scan
Shape
Slice
Split
Squeeze
Tile
Transpose
*For details, these 21 operators belong to these three models separately:*
*Bidirectional Attention Flow:*
ArgMax
Cast
CategoryMapper
Ceil
Compress
ConstantOfShape
Dropout
Gather
Hardmax
ReduceMax
ReduceSum
Scan
Shape
Slice
Squeeze
Transpose
*BERT-Squad:*
Slice
Squeeze
Shape
Identity
Gather
ReduceMean
Cast
Tile
Transpose
Split
*GPT-2:*
ConstantOfShape
Slice
Shape
Gather
ReduceMean
NonZero
Cast
Transpose
Split
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
