Skip to content

GitLab

Switch branch/tag
  1. 17 Jun, 2020 2 commits
  3. 12 Jun, 2020 2 commits
  5. 10 Jun, 2020 1 commit
  7. 05 Jun, 2020 2 commits
  9. 04 Jun, 2020 1 commit
  11. 02 Jun, 2020 1 commit
  13. 01 Jun, 2020 3 commits
  15. 29 May, 2020 1 commit
  17. 28 May, 2020 1 commit
  19. 22 May, 2020 1 commit
  21. 07 May, 2020 1 commit
  23. 04 May, 2020 1 commit
  25. 01 May, 2020 2 commits
  27. 21 Apr, 2020 1 commit
  29. 16 Apr, 2020 1 commit
  31. 14 Apr, 2020 1 commit
  33. 10 Apr, 2020 1 commit
  35. 03 Apr, 2020 1 commit
  37. 01 Apr, 2020 1 commit
  39. 26 Mar, 2020 1 commit
  41. 25 Mar, 2020 1 commit
  43. 17 Mar, 2020 3 commits
  45. 15 Mar, 2020 2 commits
  47. 13 Mar, 2020 1 commit
  49. 11 Mar, 2020 2 commits
  51. 10 Mar, 2020 1 commit
    • Ran Chen's avatar
      Save to tmp directory on non-chief workers in model_training_utils · 682d36b5
      Ran Chen authored 
      In a multi worker set up saving is done on each worker. If they're saving to the same location, e.g. GCS, there will be conflicts. With this change we save to temporary directory on non-chief workers.

Note that, there may be synchronization in saving that needs all workers to participate, so we cannot only save on one worker.

PiperOrigin-RevId: 300141152
      682d36b5
  53. 07 Mar, 2020 1 commit
  55. 06 Mar, 2020 1 commit
    • Zongwei Zhou's avatar
      Temporarily disable explicit allreduce in BERT SQuAD · 11ccb99e
      Zongwei Zhou authored 
      In BERT SQuAD, disable explicit allreduce for now to keep the original clip_by_global_norm math. With explicit allreduce, the gradients before allreduce are scaled so even if we move clip_by_global_norm before allreduce (as in TF1 and pre-TF 2.2) it will operate on scaled gradients, the math will be changed. So with explicit allreduce, it is better to move clip_by_global_norm to after allreduce.

PiperOrigin-RevId: 299278082
      11ccb99e
  57. 05 Mar, 2020 1 commit
  59. 04 Mar, 2020 1 commit