Increase tolerances in depthwise_conv_op_test.
authorJustin Lebar <jlebar@google.com>
Tue, 29 May 2018 19:05:19 +0000 (12:05 -0700)
committerTensorFlower Gardener <gardener@tensorflow.org>
Tue, 29 May 2018 19:09:00 +0000 (12:09 -0700)
PiperOrigin-RevId: 198434814

tensorflow/compiler/tests/depthwise_conv_op_test.py

index 0a0d335..03d96a2 100644 (file)
@@ -153,7 +153,7 @@ class DepthwiseConv2DTest(XLATestCase):
                   dtype=data_type).reshape(filter_in_sizes)
     with self.test_session() as sess:
       if data_type == np.float32:
-        tolerance = 1e-5
+        tolerance = 1e-4
       else:
         self.assertEqual(data_type, np.float64)
         tolerance = 1e-8
@@ -339,7 +339,7 @@ class DepthwiseConv2DTest(XLATestCase):
 
     gpu_value = _GetVal(use_xla=True)
     cpu_value = _GetVal(use_xla=False)
-    self.assertAllClose(cpu_value, gpu_value, rtol=1e-4, atol=1e-4)
+    self.assertAllClose(cpu_value, gpu_value, rtol=1e-3, atol=1e-3)
 
   def testDepthwiseConv2DInputGradCompare(self):
     for index, (input_size, filter_size, output_size, stride,