diff options
| author |  Mark Heffernan <meheff@google.com> | 2018-06-08 10:33:48 -0700 |
|---|---|---|
| committer |  TensorFlower Gardener <gardener@tensorflow.org> | 2018-06-08 10:36:51 -0700 |
| commit | 46147d8ca303e29fd15612afdb906b5220af5d3f (patch) | |
| tree | 83865335e4f4cc42ad63c609b408039bf48c8e89 | |
| parent | 0ef76693fdab2a4d1a4923444a2593f79a6b7873 (diff) | |
Increase relative error to 1e-4 on convolution_test.
convolution_test had a zero relative error bound which made it overly sensitive
to changes to the underlying computation.
PiperOrigin-RevId: 199814523
| -rw-r--r-- | tensorflow/compiler/xla/tests/convolution_test.cc | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/tensorflow/compiler/xla/tests/convolution_test.cc b/tensorflow/compiler/xla/tests/convolution_test.cc index 947959beb1..346bb3a399 100644 --- a/tensorflow/compiler/xla/tests/convolution_test.cc +++ b/tensorflow/compiler/xla/tests/convolution_test.cc @@ -47,9 +47,9 @@ class ConvolutionTest : public ClientLibraryTestBase { #if XLA_TEST_BACKEND_GPU // XLA:GPU sometimes uses FFT convolution which isn't as precise as spatial // convolution. So relax the absolute error threshold. - ErrorSpec error_spec_ = ErrorSpec(1e-2); + ErrorSpec error_spec_ = ErrorSpec(1e-2, 1e-4); #else - ErrorSpec error_spec_ = ErrorSpec(1e-4); + ErrorSpec error_spec_ = ErrorSpec(1e-4, 1e-4); #endif }; |