experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Fix an important performance regression for LSTM and GRU in tf 2.0

Browse Source 
The issue was caused by auto inline the tf function in eager context,
which cause the grappler not able to do the swap the optimization.

PiperOrigin-RevId: 251945251
This commit is contained in:
Scott Zhu 2019-06-06 15:57:19 -07:00 committed by TensorFlower Gardener
parent 9f2714e03b
commit e691be7814
1 changed files with 3 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
tensorflow/python/keras/layers/recurrent_v2.py
View File

@ -1189,6 +1189,9 @@ def _generate_defun_backend(unique_api_name, preferred_device, func):
function_attributes = { function_attributes = {
_DEFUN_API_NAME_ATTRIBUTE: unique_api_name, _DEFUN_API_NAME_ATTRIBUTE: unique_api_name,
_DEFUN_DEVICE_ATTRIBUTE: preferred_device, _DEFUN_DEVICE_ATTRIBUTE: preferred_device,
# TODO(b/133178886): The function is auto inlined in eager context, which
# make grappler fail to do the optimization. Force it to not inline here.
'_noinline': True,
} }
return function.defun_with_attributes(func=func, return function.defun_with_attributes(func=func,
attributes=function_attributes) attributes=function_attributes)