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

Keras + Distribution Strategy: Use a slightly lower tolerance for correctness test for default distribution strategy as the weights have more variance whenever the training is run on GPUs.

Browse Source 
PiperOrigin-RevId: 226117010
This commit is contained in:
Priya Gupta 2018-12-18 22:22:30 -08:00 committed by TensorFlower Gardener
parent 9a63f5b843
commit 627fd023a0
1 changed files with 4 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
tensorflow/contrib/distribute/python/keras_test.py
View File

@ -1317,17 +1317,14 @@ class TestDistributionStrategyCorrectness(test.TestCase,
@combinations.generate(strategy_and_input_combinations())
def test_correctness(self, distribution, use_numpy, use_validation_data):
# TODO(b/121224478): This test is flaky with default strategy. Remove this
# once the issue is fixed.
if isinstance(distribution, distribute_lib._DefaultDistributionStrategy): # pylint: disable=protected-access
self.skipTest('Disable the test for default strategy.')
with self.cached_session():
default_tolerance = 1e-5
tol_table = {}
if isinstance(distribution, (mirrored_strategy.MirroredStrategy,
mirrored_strategy.CoreMirroredStrategy)):
if isinstance(distribution, (
mirrored_strategy.MirroredStrategy,
mirrored_strategy.CoreMirroredStrategy,
distribute_lib._DefaultDistributionStrategy)): # pylint: disable=protected-access
# TODO(b/119257215): Weights are not exactly the same, so use larger
# tolerance for now. Predict should be related to weights.
tol_table = {