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

Assert closeness of output values instead of equality in layout optimizer tests

Browse Source 
using convolutions

Convolution output may differ across convolution algorithms and is not guaranteed to match exactly.

PiperOrigin-RevId: 205276671
This commit is contained in:
Smit Hinsu 2018-07-19 11:49:44 -07:00 committed by TensorFlower Gardener
parent 8571b4f3d2
commit 3d6c8b8aae
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/grappler/layout_optimizer_test.py
View File

@ -1340,7 +1340,7 @@ class LayoutOptimizerTest(test.TestCase):
expected_num_transposes = 2
self.assertEqual(expected_num_transposes, num_transposes)
self._assert_trans_nhwc_to_nchw('Conv2D-0', nodes)
self.assertAllEqual(output_val_ref, output_val)
self.assertAllClose(output_val_ref, output_val, atol=1e-3)
def testLoop(self):
if test.is_gpu_available(cuda_only=True):