1. 01 Apr, 2020 1 commit
  2. 31 Mar, 2020 3 commits
  3. 30 Mar, 2020 1 commit
  4. 27 Mar, 2020 3 commits
  5. 26 Mar, 2020 1 commit
  6. 25 Mar, 2020 2 commits
  7. 24 Mar, 2020 2 commits
  8. 19 Mar, 2020 3 commits
  9. 17 Mar, 2020 4 commits
  10. 12 Mar, 2020 2 commits
  11. 11 Mar, 2020 3 commits
  12. 10 Mar, 2020 2 commits
  13. 09 Mar, 2020 1 commit
  14. 07 Mar, 2020 1 commit
  15. 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
  16. 05 Mar, 2020 6 commits
  17. 04 Mar, 2020 2 commits
  18. 03 Mar, 2020 2 commits