tree 791d45a47767654d7254bd52f50dd1c91508a5fc
parent 0360adde103f3673deee624cb8e07c5b04cf99c1
author A. Unique TensorFlower <gardener@tensorflow.org> 1605745744 -0800
committer TensorFlower Gardener <gardener@tensorflow.org> 1605751166 -0800

Update gradient_checker_v2 to use a step size in the finite difference approximation that is exactly representable as a binary floating point number. This is an old trick that in some cases avoids polluting the finite difference approximation with rounding errors that cause false negatives in gradient tests.

PiperOrigin-RevId: 343180894
Change-Id: I6a48bd2500a86eadf93bb23088dd865644cabceb