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

Fix floating point exception with tf.unravel_index

Browse Source 
This PR tries to address the issue raised in 40204 where
`tf.unravel_index` caused floating point exception when any one
element is 0 in `dims`.

The issue is that `indices` in `tf.unravel_index` should make sure
it is not out of boundary compared to `dims`.

This PR fixes the issue by adding a check before hand, though Eigen.

This PR fixes 40204.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
This commit is contained in:
Yong Tang 2020-06-06 01:24:20 +00:00
parent ae76544efc
commit b0375e267c
2 changed files with 28 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
tensorflow/core/kernels/unravel_index_op.cc
View File

@ -54,6 +54,24 @@ class UnravelIndexOp : public OpKernel {
auto dims = dims_tensor.vec<Tidx>(); auto dims = dims_tensor.vec<Tidx>();
// Chek to make sure indices is not out of boundary
Eigen::Tensor<bool, 0, Eigen::RowMajor> check;
if (TensorShapeUtils::IsScalar(indices_tensor.shape())) {
auto indices = indices_tensor.scalar<Tidx>();
auto dims_prod = dims.prod();
check = (indices < dims_prod).all();
} else {
auto indices = indices_tensor.vec<Tidx>();
auto dims_prod = dims.prod()
.reshape(Eigen::array<Eigen::Index, 1>({1}))
.broadcast(
Eigen::array<Eigen::Index, 1>({indices_tensor.NumElements()}));
check = (indices < dims_prod).all();
}
OP_REQUIRES(
ctx, check(),
errors::InvalidArgument("index is out of bound as with dims"));
Eigen::array<bool, 1> reverse({true}); Eigen::array<bool, 1> reverse({true});
Tensor strides_tensor; Tensor strides_tensor;

10
tensorflow/python/kernel_tests/array_ops_test.py
View File

@ -1384,6 +1384,16 @@ class UnravelIndexTest(test_util.TensorFlowTestCase):
out_3 = array_ops.unravel_index(indices_3, dims_3) out_3 = array_ops.unravel_index(indices_3, dims_3)
self.assertAllEqual(out_3.eval(), [[3, 6, 6], [4, 5, 1]]) self.assertAllEqual(out_3.eval(), [[3, 6, 6], [4, 5, 1]])
# Test case for GitHub issue 40204.
def testUnravelIndexZeroDim(self):
with self.cached_session():
for dtype in [dtypes.int32, dtypes.int64]:
with self.assertRaisesRegexp(
errors.InvalidArgumentError, "index is out of bound as with dims"):
indices = constant_op.constant([2, 5, 7], dtype=dtype)
dims = constant_op.constant([3, 0], dtype=dtype)
self.evaluate(array_ops.unravel_index(indices=indices, dims=dims))
class GuaranteeConstOpTest(test_util.TensorFlowTestCase): class GuaranteeConstOpTest(test_util.TensorFlowTestCase):