Commit Graph

  • b43468bb67 Reduce layer-drop prob Daniel Povey 2022-03-26 19:36:33 +08:00
  • 8a38d9a855 Fix/patch how fix_random_seed() is imported. Daniel Povey 2022-03-26 15:43:47 +08:00
  • 26a1730392 Add random-number-setting function in dataloader Daniel Povey 2022-03-26 14:46:27 +08:00
  • 0e694739f2 Fix test mode with random layer dropout Daniel Povey 2022-03-25 23:28:52 +08:00
  • 609f76675f do a fix luomingshuang 2022-03-25 20:56:13 +08:00
  • d2ed3dfc90 Fix bug Daniel Povey 2022-03-25 20:35:11 +08:00
  • 4b650e9f01 Make warmup work by scaling layer contributions; leave residual layer-drop Daniel Povey 2022-03-25 20:34:33 +08:00
  • 5553a086d0 do a change for RESULTS.md luomingshuang 2022-03-25 19:32:23 +08:00
  • 33fbde8d97 do a change for RESULTS.md luomingshuang 2022-03-25 19:31:07 +08:00
  • cc68bc256b add fast beam search for decoding luomingshuang 2022-03-25 19:26:58 +08:00
  • aecb6dce71 Randomly combining output from different transformer encoder layers. Fangjun Kuang 2022-03-25 17:39:57 +08:00
  • 12de88043a Randomly combining output from encoder layers. Fangjun Kuang 2022-03-25 17:13:22 +08:00
  • 1f548548d2 Simplify the warmup code; max_abs 10->6 Daniel Povey 2022-03-24 15:06:06 +08:00
  • aab72bc2a5 Add changes from master to decode.py, train.py Daniel Povey 2022-03-24 13:10:54 +08:00
  • 5d9dae3064 Merge changes from master Daniel Povey 2022-03-24 12:59:36 +08:00
  • 2da510cb8b Merge branch 'master' of https://github.com/k2-fsa/icefall into spgi Desh Raj 2022-03-23 15:44:23 -04:00
  • 395a3f952b
    Batch decoding for models trained with optimized_transducer (#267) Fangjun Kuang 2022-03-23 19:11:34 +08:00
  • d21ac737ad Minor fixes. Fangjun Kuang 2022-03-23 19:11:01 +08:00
  • 055406a67e Minor fixes. Fangjun Kuang 2022-03-23 17:35:39 +08:00
  • 127155b37d Minor fixes Fangjun Kuang 2022-03-23 17:23:22 +08:00
  • ced16beab0 Minor fixes. Fangjun Kuang 2022-03-23 16:23:29 +08:00
  • 8cc5cd81b3 Add modified beam search in batch mode. Fangjun Kuang 2022-03-23 15:44:33 +08:00
  • 7fa5860073 Add greedy search in batch mode. Fangjun Kuang 2022-03-23 11:43:26 +08:00
  • aa71eaaac7 Refactoring. Fangjun Kuang 2022-03-23 11:10:01 +08:00
  • 3ae7265737
    More fixes to the checkpoint code. (#266) Fangjun Kuang 2022-03-23 14:37:54 +08:00
  • 20a965e8a9 More fixes to the checkpoint code. Fangjun Kuang 2022-03-23 14:17:53 +08:00
  • 1ce4349c17
    Merge adb54aea91abe211b19ec75eeb422b15a3867405 into 6a091da0b0543befb0492848d3583700c274d111 Fangjun Kuang 2022-03-23 12:52:43 +08:00
  • 6a091da0b0
    Minor fixes for saving checkpoints. (#265) Fangjun Kuang 2022-03-23 12:22:05 +08:00
  • 4fe0c0dca2 Fix loading checkpoints saved by previous code. Fangjun Kuang 2022-03-23 12:21:11 +08:00
  • ed2d2932ef Minor fixes for saving checkpoints. Fangjun Kuang 2022-03-23 12:02:06 +08:00
  • d38a1480b9 Minor fixes beam_search Fangjun Kuang 2022-03-23 11:35:24 +08:00
  • 9a8aa1f54a Change how warmup works. Daniel Povey 2022-03-22 15:36:20 +08:00
  • 8c7995d493
    Support modified beam search in batch mode. (#264) Fangjun Kuang 2022-03-22 15:14:04 +08:00
  • 4808dd3fdb Minor fixes. Fangjun Kuang 2022-03-22 14:53:17 +08:00
  • 4187ad0d06 Update k2 versions in GitHub CI. Fangjun Kuang 2022-03-22 14:36:40 +08:00
  • c7bd7e5db7 Typo fixes. Fangjun Kuang 2022-03-22 14:34:09 +08:00
  • cef6348703 Change max-abs from 6 to 10 Daniel Povey 2022-03-22 13:50:54 +08:00
  • 4004ca81d8 Increase warm_step (and valid_interval) Daniel Povey 2022-03-22 13:32:24 +08:00
  • 708f31247f Support modified beam search in batch mode. Fangjun Kuang 2022-03-22 13:00:24 +08:00
  • b82a505dfc Reduce initial pruned_loss scale from 0.01 to 0.0 Daniel Povey 2022-03-22 12:30:48 +08:00
  • d5c78a2238
    Implement greedy search in batch mode for transducer decoding. (#262) Fangjun Kuang 2022-03-22 10:32:22 +08:00
  • b7e84d5d77 Whitespace fix Daniel Povey 2022-03-21 23:59:53 +08:00
  • 2eef001d39 Fix balancer code Daniel Povey 2022-03-21 23:59:26 +08:00
  • 55e3019c5d
    Update .flake8 Wang, Guanbo 2022-03-21 11:04:32 -04:00
  • 3e96feae37 Implement greedy search in batch mode for transducer decoding. Fangjun Kuang 2022-03-21 22:16:51 +08:00
  • 11a04c50ae Change 0.025,0.05 to 0.01 in initializations Daniel Povey 2022-03-21 21:29:24 +08:00
  • 05e30d0c46 Add max-abs=6, debugged version Daniel Povey 2022-03-21 21:15:00 +08:00
  • d844655a63 Merge remote-tracking branch 'upstream/master' into rework2e_ws30k_16k_0.025_ma6 Daniel Povey 2022-03-21 21:13:25 +08:00
  • ccbf8ba086 Incorporate changes from master into pruned_transducer_stateless2. Daniel Povey 2022-03-21 16:51:48 +08:00
  • ff855ff821 update README.md luomingshuang 2022-03-21 21:10:04 +08:00
  • 26146e8282 update README.md luomingshuang 2022-03-21 21:07:54 +08:00
  • eeddeaeccb update tedlium3-pruned-transducer-stateless-codes luomingshuang 2022-03-21 21:04:23 +08:00
  • 7896baea14 Add modified-beam-search and fast-beam-search pkufool 2022-03-21 19:31:51 +08:00
  • 7a3e88d2d3 Merge branch 'master' into pruned_aishell pkufool 2022-03-21 17:30:57 +08:00
  • b2b4d9e0b6
    Add fast beam search decoding (#250) Wei Kang 2022-03-21 16:22:25 +08:00
  • 05b5e78d8f Add norm+balancer to VggSubsampling Daniel Povey 2022-03-21 15:55:11 +08:00
  • 8f60e347fb
    Merge branch 'master' into update-diagnostics Mingshuang Luo 2022-03-21 14:37:28 +08:00
  • 2ffe72af81 fix a rare error luomingshuang 2022-03-21 14:28:53 +08:00
  • c3993a53d9
    Merge branch 'k2-fsa:master' into gigaspeech_recipe Wang, Guanbo 2022-03-21 01:49:39 -04:00
  • ae564f91e6
    Periodically saving checkpoint after processing given number of batches (#259) Fangjun Kuang 2022-03-20 23:51:33 +08:00
  • 3dc34a0a82 Fixes after review. Fangjun Kuang 2022-03-20 23:30:07 +08:00
  • a1189615f3 Save checkpoints periodically. Fangjun Kuang 2022-03-20 22:40:14 +08:00
  • 910e6c9306
    Minor fixes to tedlimu3 to make ./prepare.sh working. (#258) Fangjun Kuang 2022-03-20 20:26:03 +08:00
  • 1a9e8a5718 WIP: Periodically saving checkpoint after processing given number of batches. Fangjun Kuang 2022-03-20 19:42:20 +08:00
  • a0db9cda93 Minor fixes to tedlimu3 to make ./prepare.sh working. Fangjun Kuang 2022-03-20 16:17:36 +08:00
  • 0ee2404ff0 Remove logging code that broke with newer Lhotse; fix bug with pruned_loss Daniel Povey 2022-03-19 14:01:45 +08:00
  • 77e2c4a28a
    Merge 054e2399b9d21cbd1aac6186f80997f0eef2104f into ad28c8c5ebef4cef33b99482469c83e7b36ff07d Fangjun Kuang 2022-03-18 17:28:07 +01:00
  • 8cff994cd7 Set also scale for embedding to 0.025. Daniel Povey 2022-03-18 21:30:05 +08:00
  • 188eada7ac Change initial std from 0.05 to 0.025. Daniel Povey 2022-03-18 21:28:34 +08:00
  • c9f1aeb7d1 Fix bug with import Daniel Povey 2022-03-18 16:40:24 +08:00
  • 2dfcd8f117 Double warm_step Daniel Povey 2022-03-18 16:38:36 +08:00
  • ba3611cefd Cosmetic changes to swish Daniel Povey 2022-03-18 16:35:48 +08:00
  • 6769087d70 Remove scale_speed, make swish deriv more efficient. Daniel Povey 2022-03-18 16:31:25 +08:00
  • ad28c8c5eb
    Tedlium3 transducer stateless (#233) Mingshuang Luo 2022-03-18 11:39:06 +08:00
  • 10ebd9f388 do a small change for prepare.sh luomingshuang 2022-03-18 11:32:33 +08:00
  • b68ddb03f4 do a change luomingshuang 2022-03-18 11:05:32 +08:00
  • 2a5be552a7 Fix comments pkufool 2022-03-18 10:46:10 +08:00
  • 089b8178f0 Fix comments pkufool 2022-03-18 10:45:47 +08:00
  • 6772a3cf84 update codes luomingshuang 2022-03-18 10:29:15 +08:00
  • cbe6b175d1 Reduce warmup scale on pruned loss form 0.1 to 0.01. Daniel Povey 2022-03-17 16:46:59 +08:00
  • acc0eda5b0 Scale down pruned loss in warmup mode Daniel Povey 2022-03-17 16:09:35 +08:00
  • 13db33ffa2 Fix diagnostics-getting code Daniel Povey 2022-03-17 15:53:53 +08:00
  • 11bea4513e Add remaining files in pruned_transducer_stateless2 Daniel Povey 2022-03-17 11:17:52 +08:00
  • b2a2b1d387 add decoding Desh Raj 2022-03-16 13:25:35 -04:00
  • e3ad8f63e7 update decode.py file type Daniel Povey 2022-03-16 22:22:10 +08:00
  • cc8e4412f7 Add more files.. Daniel Povey 2022-03-16 22:16:40 +08:00
  • 1f3a15f3c4 Start adding some files.. Daniel Povey 2022-03-16 22:14:30 +08:00
  • 87c92efbfe Changes from upstream/master Daniel Povey 2022-03-16 21:49:15 +08:00
  • 518ec6414a
    Update diagnostics.py (#254) Mingshuang Luo 2022-03-16 20:17:45 +08:00
  • e838c192ef Cosmetic changes/renaming things Daniel Povey 2022-03-16 19:27:45 +08:00
  • dfc75752c4 Remove some dead code. Daniel Povey 2022-03-16 18:06:01 +08:00
  • c82db4184a Remove xscale from pos_embedding Daniel Povey 2022-03-16 15:50:11 +08:00
  • 6561743d7b bug fix re sqrt Daniel Povey 2022-03-16 14:55:17 +08:00
  • 0e9cad3f1f Modifying initialization from normal->uniform; add initial_scale when initializing Daniel Povey 2022-03-16 14:42:53 +08:00
  • 00be56c7a0 Remove dead code Daniel Povey 2022-03-16 12:49:00 +08:00
  • a783b96467 Fix typo Daniel Povey 2022-03-16 12:43:44 +08:00
  • 633213424d Rework of initialization Daniel Povey 2022-03-16 12:42:59 +08:00
  • dbfe8fbb1a Minor fixes pkufool 2022-03-16 11:53:03 +08:00
  • 1331199530 Merge branch 'specaugmod_baseline' into randcombine1_expscale3_rework2c_maxabs1000_maxp0.95_noexp_convderiv2warmup_scale_0mean Daniel Povey 2022-03-15 23:47:03 +08:00
  • 261d7602a7 Draft of 0mean changes.. Daniel Povey 2022-03-15 23:46:53 +08:00