Commits · af7a22507e1f7d8a429678a4fd04783b466161ff · KMSCAKKSCFKA AKFACAMADCAS / tensorflow-federated

This project is mirrored from https://github.com/tensorflow/federated. Pull mirroring failed 1 year ago.
Repository mirroring has been paused due to too many failed attempts. It can be resumed by a project maintainer or owner.
Last successful update 1 year ago.

Oct 05, 2019
- Factor checkpoint serial number extraction out to module-level. · af7a2250
  A. Unique TensorFlower authored 5 years ago
  
  Use the serial number pattern to check target directory naming before saving a checkpoint. PiperOrigin-RevId: 272911215
  af7a2250
Oct 04, 2019
- Take advantage of "key" kwarg when computing the max checkpoint. · 97dd1496
  A. Unique TensorFlower authored 5 years ago
  
  PiperOrigin-RevId: 272474618
  97dd1496
Sep 17, 2019
- Update checkpoint_utils to make atomic writes. · 409292bf
  Zachary Garrett authored 5 years ago
  
  PiperOrigin-RevId: 269437287
  409292bf
- Fix code to be TF 2.0 compatible. · 435de5fc
  Zachary Garrett authored 5 years ago
  
  PiperOrigin-RevId: 268942830
  435de5fc
Sep 04, 2019

Revert checkpoint_utils behavior to find the latest checkpoint based on filename. · 06dcd833

Zachary Garrett authored 5 years ago

Depending on how fast the checkpoints are written, the mtimes may sometimes be the same
(not enough resolution); unclear how to tie-break and causes issues.

PiperOrigin-RevId: 267155543

06dcd833

Simplify checkpoint utils and remove numpy dependency. · ca179ac0

Zachary Garrett authored 5 years ago

- Allow user to specify the checkpoint prefix.
- Depend on the checkpoints file mtime for determining 'latest'
- Rewrite unittests to not rely on `tff.learning`

PiperOrigin-RevId: 267023625

ca179ac0

Aug 23, 2019

Adds write_state and read_state in research/utils. · d80d6923

Zhouyuan Huo authored 5 years ago

This is used to resume the training of federated learning after the job is stopped locally or preempted on borg.

PiperOrigin-RevId: 264915177

d80d6923