mirrors/rocm_jax
1
0
Fork 0
mirror of https://github.com/ROCm/jax.git synced 2025-04-25 00:56:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

[jax2tf] Update limitations with information about potentially

Browse Source 
equally valid but different expected results.
This commit is contained in:
Benjamin Chetioui 2020-12-16 11:05:15 +01:00
parent 4bba27a446
commit 1dc55aabdf
2 changed files with 40 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
jax/experimental/jax2tf/g3doc/primitives_with_limited_support.md
View File

@ -60,6 +60,26 @@ Additionally, some primitives have numerical differences between JAX and TF in s
for completion purposes. It is not expected that this
operation will ever have a use case.
* `acos`, `asin`, `atan`, and their hyperbolic counterparts
`acosh`, `asinh`, `atanh` return equally valid but possibly
different results when the parameter has a complex dtype.
* `min` and `max` return different values when one of the elements
being compared is NaN. JAX always returns NaN, while TF returns
the value NaN is compared with.
* `cholesky` returns potentially different values in the strictly
upper triangular part of the result. This does not matter for
correctness, because this part of the matrix is not considered in
the result.
* `eig` and `eigh` return eigenvalues and eigenvectors in a
potentially different order. The eigenvectors may also be
different, but equally valid.
* `lu` may return different but equally valid results when the LU
decomposition is not unique.
## Updating the documentation
This file is automatically generated by running the jax2tf tests in

20
jax/experimental/jax2tf/g3doc/primitives_with_limited_support.md.template
View File

@ -60,6 +60,26 @@ Additionally, some primitives have numerical differences between JAX and TF in s
for completion purposes. It is not expected that this
operation will ever have a use case.
* `acos`, `asin`, `atan`, and their hyperbolic counterparts
`acosh`, `asinh`, `atanh` return equally valid but possibly
different results when the parameter has a complex dtype.
* `min` and `max` return different values when one of the elements
being compared is NaN. JAX always returns NaN, while TF returns
the value NaN is compared with.
* `cholesky` returns potentially different values in the strictly
upper triangular part of the result. This does not matter for
correctness, because this part of the matrix is not considered in
the result.
* `eig` and `eigh` return eigenvalues and eigenvectors in a
potentially different order. The eigenvectors may also be
different, but equally valid.
* `lu` may return different but equally valid results when the LU
decomposition is not unique.
## Updating the documentation
This file is automatically generated by running the jax2tf tests in