Releases: NVIDIA/framework-reproducibility
Releases · NVIDIA/framework-reproducibility
Version 0.6.0
This is a re-release of version 0.5.0, which was accidentally released with stale files.
Version 0.5.0
In Seeder, add the ability to reseed generators with either
- a different seed for each worker (the existing functionality) or
- the same seed for each worker (added functionality).
Version 0.4.0
Enhanced Functionality
- Add the Seeder tool for variance reduction. This is an experimental feature.
- Rename the distribution from
tensorflow-determinism
toframework-reproducibility
and rename the package fromtfdeterminism
tofwr13y
. - Add
fwr13y.d9m.tensorflow.enable_determinism
, which makes a best-effort to enable determinism in whichever version of TensorFlow is being used. - Add a script to find commits in the TensorFlow repo related to determinism.
fwr13y.d9m.tensorflow.patch
throws more specific exceptions.
Enhanced Testing / Higher Quality
- Test patched determinism over a wider range of stock TensorFlow and NGC TensorFlow versions.
Version 0.3.0
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.
Version 0.2.0
New Functionality
- Add patch availability on TensorFlow version 1.15
- Print the version of tensorflow-determinism when patch is applied
Enhanced Testing / Higher Quality
- Test that patch will throw exception on non-supported versions of TF
- Test that patch will throw exception in NGC containers
- Test that patch works in Python 3
- Test that package will install when TensorFlow is not yet installed
Developed by Duncan Riach with thanks to Nathan Luehr for review.
Version 0.1.0
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.