Providing reproducibility in deep learning frameworks
This is a re-release of version 0.5.0, which was accidentally released with stale files.
In Seeder, add the ability to reseed generators with either
tensorflow-determinism
to framework-reproducibility
and rename the package from tfdeterminism
to fwr13y
.fwr13y.d9m.tensorflow.enable_determinism
, which makes a best-effort to enable determinism in whichever version of TensorFlow is being used.fwr13y.d9m.tensorflow.patch
throws more specific exceptions.Add patch availability for stock TensorFlow version 2.0, and test in eager mode.
Developed by Duncan Riach with thanks to Nathan Luehr for review.
This release includes a patch for standard TF 1.14.0 that enables most deep learning TF models to train deterministically on GPUs. GPU-determinism support in the NVIDIA NGC TF containers is also described.
Developed by Duncan Riach with thanks to Nathan Luehr for review.
Developed by Duncan Riach with thanks to Nathan Luehr for review.