Commit Graph

  • f1cf817cf4 change for asr_datamodule.py luomingshuang 2022-03-07 16:46:26 +08:00
  • b67c57de92 change for asr_datamodule.py luomingshuang 2022-03-07 16:32:52 +08:00
  • 2aca0d536c Fix more typos. Fangjun Kuang 2022-03-07 16:29:13 +08:00
  • fb63ed627d Fix typos. Fangjun Kuang 2022-03-07 16:19:06 +08:00
  • 5df6040df0 Update README to include force alignment information. Fangjun Kuang 2022-03-07 16:03:39 +08:00
  • 93aaa14cda change asr_datamodule.py luomingshuang 2022-03-07 16:03:09 +08:00
  • c26a7e4dc4 change asr_datamodule.py luomingshuang 2022-03-07 15:59:47 +08:00
  • 75936a5fae Compute word starting time from framewise token alignment. Fangjun Kuang 2022-03-07 15:16:54 +08:00
  • d50e7734a6 Add more documentation. Fangjun Kuang 2022-03-07 12:57:11 +08:00
  • e2ace9d545 Replace norm on input layer with scale of 0.1. Daniel Povey 2022-03-07 11:24:04 +08:00
  • 7f9e426878 fix comment luomingshuang 2022-03-07 10:55:06 +08:00
  • 2cbba6901e fix comments luomingshuang 2022-03-07 10:13:03 +08:00
  • 6bcfa6225f Add force alignment for stateless transducer. Fangjun Kuang 2022-03-06 23:14:03 +08:00
  • a37d98463a Restore ConvolutionModule to state before changes; change all Swish,Swish(Swish) to SwishOffset. Daniel Povey 2022-03-06 11:55:02 +08:00
  • 86b93ff30d
    Merge b79c733cd67ee29119a6d0f77cef02f0e8836545 into 1603744469d167d848e074f2ea98c587153205fa Fangjun Kuang 2022-03-06 11:37:28 +08:00
  • 8a8b81cd18 Replace relu with swish-squared. Daniel Povey 2022-03-05 22:21:42 +08:00
  • 1603744469
    Refactor conformer. (#237) Fangjun Kuang 2022-03-05 19:26:06 +08:00
  • b9c3892da1 Refactor conformer. Fangjun Kuang 2022-03-05 18:51:35 +08:00
  • 5f2c0a09b7 Convert swish nonlinearities to ReLU Daniel Povey 2022-03-05 16:28:24 +08:00
  • 0cd14ae739 Fix exp dir Daniel Povey 2022-03-05 12:17:09 +08:00
  • 65b09dd5f2 Double the threshold in brelu; slightly increase max_factor. Daniel Povey 2022-03-05 00:07:14 +08:00
  • 74f2b163de Merge diagnostics improvement Daniel Povey 2022-03-04 23:15:47 +08:00
  • 6252282fd0 Add deriv-balancing code Daniel Povey 2022-03-04 20:19:11 +08:00
  • eb3ed54202 Reduce scale from 50 to 20 Daniel Povey 2022-03-04 15:56:45 +08:00
  • 9cc5999829 Fix duplicate Swish; replace norm+swish with swish+exp-scale in convolution module Daniel Povey 2022-03-04 15:50:51 +08:00
  • ad62981765
    Add diagnostics (#230) yaozengwei 2022-03-04 15:38:23 +08:00
  • c68adbbf85 Corrected the time in copyright information. yaozengwei 2022-03-04 15:13:14 +08:00
  • dab2b77210 Add copyright information. yaozengwei 2022-03-04 15:08:35 +08:00
  • 7e88999641 Increase scale from 20 to 50. Daniel Povey 2022-03-04 14:31:29 +08:00
  • 3207bd98a9 Increase scale on Scale from 4 to 20 Daniel Povey 2022-03-04 13:16:40 +08:00
  • 503f8d521c Fix bug in diagnostics Daniel Povey 2022-03-04 13:08:56 +08:00
  • 3d9ddc2016 Fix backprop bug Daniel Povey 2022-03-04 12:29:44 +08:00
  • 2f0fbf430c
    Remove duplicate files. (#236) Fangjun Kuang 2022-03-04 11:56:31 +08:00
  • 4a7f73e704 Remove duplicate files. Fangjun Kuang 2022-03-04 11:23:39 +08:00
  • cd216f50b6 Add import Daniel Povey 2022-03-04 11:03:01 +08:00
  • bc6c720e25 Combine ExpScale and swish for memory reduction Daniel Povey 2022-03-04 10:52:05 +08:00
  • 23b3aa233c Double learning rate of exp-scale units Daniel Povey 2022-03-04 00:42:37 +08:00
  • 5c177fc52b pelu_base->expscale, add 2xExpScale in subsampling, and in feedforward units. Daniel Povey 2022-03-03 23:52:03 +08:00
  • 3ec219dfa0
    Add stateless transducer tutorial. (#235) Fangjun Kuang 2022-03-03 22:33:47 +08:00
  • d1c0c19ab3 Minor fixes. Fangjun Kuang 2022-03-03 22:33:13 +08:00
  • dfe0bc43d7 Add more doc. Fangjun Kuang 2022-03-03 17:27:42 +08:00
  • 973149d3bc fix spelling errors luomingshuang 2022-03-03 17:21:47 +08:00
  • 8be385f3bd Update docs of arguments. yaozengwei 2022-03-03 17:13:33 +08:00
  • 35812fd6de do some changes luomingshuang 2022-03-03 16:28:51 +08:00
  • 87b4619f12 Update docs of arguments, and remove stats_types() function in TensorDiagnosticOptions object. yaozengwei 2022-03-03 15:35:12 +08:00
  • 194a4e6864 change for export.py luomingshuang 2022-03-02 21:48:58 +08:00
  • 85cfd412e2 replace some py files with symlink luomingshuang 2022-03-02 21:28:46 +08:00
  • 3fb559d2f0 Add baseline for the PeLU expt, keeping only the small normalization-related changes. Daniel Povey 2022-03-02 18:27:08 +08:00
  • 334f8bb906 WIP: Add stateless transducer tutorial. Fangjun Kuang 2022-03-02 18:01:10 +08:00
  • 1ff6196c44
    Fix joiner (#234) Fangjun Kuang 2022-03-02 16:41:14 +08:00
  • 9ed7d55a84 Small bug fixes/imports Daniel Povey 2022-03-02 16:34:55 +08:00
  • 9d1b4ae046 Add pelu to this good-performing setup.. Daniel Povey 2022-03-02 16:33:27 +08:00
  • 22bc87d931 Remove duplicate files. Fangjun Kuang 2022-03-02 16:16:24 +08:00
  • fbeca54fa9 update RESULTS.md luomingshuang 2022-03-02 16:16:49 +08:00
  • 9b62bfb062 Add tests for Joiner Fangjun Kuang 2022-03-02 16:13:52 +08:00
  • 50d2281524
    Add modified transducer loss for AIShell dataset (#219) Fangjun Kuang 2022-03-02 16:02:38 +08:00
  • 313670e26a Use symlinks to avoid copies. Fangjun Kuang 2022-03-02 15:47:50 +08:00
  • 73d2cb7c1e add transducer stateless files luomingshuang 2022-03-02 15:16:53 +08:00
  • 94810ced45 add md files and prepare.sh luomingshuang 2022-03-02 15:12:10 +08:00
  • 23dc6a0a84 add files luomingshuang 2022-03-02 14:45:17 +08:00
  • c7ada15929 Update README. Fangjun Kuang 2022-03-02 14:09:13 +08:00
  • 9ee8a4defa Update results. Fangjun Kuang 2022-03-02 14:07:38 +08:00
  • 6f774b0e28 add tedlium3 transducer-stateless luomingshuang 2022-03-02 11:46:51 +08:00
  • acbe78ccc4 add tedlium3 transducer-stateless luomingshuang 2022-03-02 11:46:19 +08:00
  • 131f087dcb Update results. Fangjun Kuang 2022-03-02 11:33:11 +08:00
  • c80dae7b07 Update CI to install pretrained models with aishell. Fangjun Kuang 2022-03-01 18:06:29 +08:00
  • e6e0accfa5 Merge remote-tracking branch 'dan/master' into aishell-modified-transducer Fangjun Kuang 2022-03-01 18:00:14 +08:00
  • a5c3dfea1e Update export.py and pretrained.py Fangjun Kuang 2022-03-01 17:59:56 +08:00
  • 05cb297858
    Update result for full libri + GigaSpeech using transducer_stateless. (#231) Fangjun Kuang 2022-03-01 17:01:46 +08:00
  • cc2628f085 Update result for full libri + GigaSpeech using transducer_stateless. Fangjun Kuang 2022-03-01 16:27:34 +08:00
  • 72f838dee1
    Update results for transducer_stateless after training for more epochs. (#207) Fangjun Kuang 2022-03-01 16:35:02 +08:00
  • 828a23daad Remove the diagnostics code in the local dir yaozengwei 2022-03-01 01:14:29 +08:00
  • 0b2fa43009 Move diagnostics code from local dir to the shared icefall dir yaozengwei 2022-03-01 00:45:33 +08:00
  • c660c71324 Adding diagnostics code... Daniel Povey 2022-02-27 13:44:43 +08:00
  • 2ff520c800 Improvements to diagnostics (RE those with 1 dim Daniel Povey 2022-02-28 12:15:56 +08:00
  • c1063def95 First version of rand-combine iterated-training-like idea. Daniel Povey 2022-02-27 17:34:58 +08:00
  • 63d8d935d4 Refactor/simplify ConformerEncoder Daniel Povey 2022-02-27 13:56:15 +08:00
  • 581786a6d3 Adding diagnostics code... Daniel Povey 2022-02-27 13:44:43 +08:00
  • ab639ed7f9 Add extra data in transducer training. Fangjun Kuang 2022-02-26 17:50:43 +08:00
  • 008bad624c
    Merge a533143d2f32dfdacfb8036b18f12aeb175638c7 into ac7c2d84bc44dd83a5653f68677ae1ef16551eea Mingshuang Luo 2022-02-25 04:15:47 +00:00
  • a533143d2f fix conflicts Mingshuang Luo 2022-02-25 12:15:03 +08:00
  • 960100a4a8
    Merge branch 'k2-fsa:master' into master Mingshuang Luo 2022-02-25 12:06:49 +08:00
  • 8836cf0d1f update .flake8 Mingshuang Luo 2022-02-25 12:01:19 +08:00
  • 2f293ce5d3 update RESULTS.md Mingshuang Luo 2022-02-25 11:46:47 +08:00
  • b5f408689e update training data maximum duration Mingshuang Luo 2022-02-25 11:40:57 +08:00
  • e752c0440d Update RESULTS.md Mingshuang Luo 2022-02-24 19:59:25 +08:00
  • 536ad2252e Update tedlium3 transducer stateless Mingshuang Luo 2022-02-24 19:57:36 +08:00
  • 9011ebbf9d
    Merge f8065fb2791b9713d8b5b8af0b5b468bce5567a2 into ac7c2d84bc44dd83a5653f68677ae1ef16551eea Mingshuang Luo 2022-02-23 10:13:13 +08:00
  • ac7c2d84bc
    minor fix for aishell recipe (#223) PF Luo 2022-02-23 08:33:20 +08:00
  • 4319a187b3 Zero out the gradient of decoder/joiner for auxiliary losses. Fangjun Kuang 2022-02-22 19:44:01 +08:00
  • 76632bddfe Minor fixes. Fangjun Kuang 2022-02-22 17:37:14 +08:00
  • 8b4c571824 Add auxiliary losses. Fangjun Kuang 2022-02-22 16:15:00 +08:00
  • 2450b33f54 minor fixs for aishell PingFeng Luo 2022-02-22 15:42:53 +08:00
  • e558be1cb7 just remove unnecessary torch.sum PingFeng Luo 2022-02-22 14:41:50 +08:00
  • 3b34c7591a
    Merge d412dbb2f0a47f1b7385322fd0d6251e678554c8 into 2332ba312d7ce72f08c7bac1e3312f7e3dd722dc Mingshuang Luo 2022-02-22 11:52:39 +08:00
  • 5e4b1e01fe Get the input for the auxiliary branch. Fangjun Kuang 2022-02-21 19:56:45 +08:00
  • 218d79aba7 Copy files. Fangjun Kuang 2022-02-21 19:32:26 +08:00
  • 2332ba312d
    Begin to use multiple datasets in training (#213) Fangjun Kuang 2022-02-21 15:27:27 +08:00
  • 1c35ae1dba
    Reset seed at the beginning of each epoch. (#221) Fangjun Kuang 2022-02-21 15:16:39 +08:00
  • aadd7cae42 Minor fixes. Fangjun Kuang 2022-02-21 15:11:46 +08:00