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

Mirrored strategy and default strategy can result in different layouts in XLA that can result in different accumulation orders for reduction operations. As a result, relax the precision of the keras correctness tests.

Browse Source 
PiperOrigin-RevId: 327438391
Change-Id: I1b8a8e119cff89bce991466e8289568e96cc6d31
This commit is contained in:
A. Unique TensorFlower 2020-08-19 08:34:22 -07:00 committed by TensorFlower Gardener
parent 4a896124e3
commit e5f12a0ff5
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/python/keras/distribute/keras_correctness_test_base.py
View File

@ -308,7 +308,7 @@ def compare_results(results_with_ds,
default_tolerance = 1e-3 default_tolerance = 1e-3
relaxed_tolerance = 1e-3 relaxed_tolerance = 1e-3
else: else:
default_tolerance = 1e-5 default_tolerance = 4e-5
relaxed_tolerance = 1e-4 relaxed_tolerance = 1e-4
def _get_compare_result_tolerance(key): def _get_compare_result_tolerance(key):