Commit Graph

  • 3d1564d735 add delay penalty yaozengwei 2022-09-19 23:16:22 +08:00
  • b628fae348 Add Dockerfiles Teo Wen Shen 2022-09-19 14:53:01 +00:00
  • 8b45ac0438
    Update README.md Teo Wen Shen 2022-09-19 22:26:41 +08:00
  • 8b9211ec62 init files yaozengwei 2022-09-19 20:41:57 +08:00
  • 6fdc9d53a4 Add missing files Fangjun Kuang 2022-09-19 20:34:15 +08:00
  • 104dce59da Merge remote-tracking branch 'k2-fsa/master' yaozengwei 2022-09-19 20:30:13 +08:00
  • 8776c56491 support exporting to ncnn format via PNNX Fangjun Kuang 2022-09-19 20:14:54 +08:00
  • 3d72a65de8 Implement max-eig-proportion.. Daniel Povey 2022-09-19 10:26:37 +08:00
  • 5f27cbdb44 Merge branch 'scaled_adam_exp4_max_var_per_eig' into scaled_adam_exp7 Daniel Povey 2022-09-18 21:23:59 +08:00
  • 0f567e27a5 Add max_var_per_eig in self-attn Daniel Povey 2022-09-18 21:22:01 +08:00
  • eb77fa7aaa Restore min_positive,max_positive limits on linear_pos projection Daniel Povey 2022-09-18 14:38:30 +08:00
  • 69404f61ef Use scalar_lr_scale for scalars as well as sizes. Daniel Povey 2022-09-18 14:12:27 +08:00
  • 76031a7c1d Loosen some limits of activation balancers Daniel Povey 2022-09-18 13:58:19 +08:00
  • 3122637266 Use ScaledLinear where I previously had StructuredLinear Daniel Povey 2022-09-17 13:18:58 +08:00
  • 4a2b940321 Remove StructuredLinear,StructuredConv1d Daniel Povey 2022-09-17 13:14:08 +08:00
  • 1a184596b6 A little code refactoring Daniel Povey 2022-09-16 20:56:21 +08:00
  • 97b3fc53aa
    Add LSTM for the multi-dataset setup. (#558) Fangjun Kuang 2022-09-16 18:40:25 +08:00
  • bb1bee4a7b Improve how quartiles are printed Daniel Povey 2022-09-16 17:30:03 +08:00
  • 5f55f80fbb Configure train.py with clipping_scale=2.0 Daniel Povey 2022-09-16 17:19:52 +08:00
  • 8298333bd2 Implement gradient clipping. Daniel Povey 2022-09-16 16:52:46 +08:00
  • 8f876b3f54 Remove batching from ScaledAdam, in preparation to add gradient norm clipping Daniel Povey 2022-09-16 15:42:56 +08:00
  • 3b450c2682 Bug fix in train.py, fix optimzier name Daniel Povey 2022-09-16 14:10:42 +08:00
  • 257c961b66 1st attempt at scaled_adam Daniel Povey 2022-09-16 13:50:15 +08:00
  • 96dbec7807
    Merge 7d5f491e3ac6a26e76fdc228d8dbc69f391c9183 into 145c44f71095f174a6d994a479074f105a4fb6ad Zengwei Yao 2022-09-16 02:02:06 +09:00
  • 1076d0f7cc
    Merge fb2a866e1fd3a7b605d82cf7743a5b0926062ab2 into 145c44f71095f174a6d994a479074f105a4fb6ad Fangjun Kuang 2022-09-16 02:00:50 +09:00
  • a5747c2730 filter with soft mask yaozengwei 2022-09-15 11:32:04 +08:00
  • fd19a573b7
    Merge branch 'k2-fsa:master' into lightweight Tiance Wang 2022-09-13 17:10:56 +08:00
  • e71b541a21 fix typing and refactor yaozengwei 2022-09-13 13:50:01 +08:00
  • 145c44f710
    Use modified ctc topo when vocab size is > 500 (#568) Fangjun Kuang 2022-09-13 10:59:27 +08:00
  • 102f77ba1f Use modified ctc topo when vocab size is > 500 Fangjun Kuang 2022-09-13 10:51:35 +08:00
  • af2fcf25cb apply gradient filter in LSTM module, to filter both input and params yaozengwei 2022-09-13 10:17:12 +08:00
  • be65fde9f7
    Merge 6edb4d3806c98b172dea8d2ef5b2514b2b13ac32 into 9e24642faf985e50425b78caef5517c87c0c7efc Zengwei Yao 2022-09-11 15:58:23 -04:00
  • 9e24642faf
    Modified prepare_transcripts.py and preprare_lexicon.py of tedlium3 recipe (#567) shcxlee 2022-09-09 21:32:49 -05:00
  • 85c71685fc
    Update prepare_transcripts.py shcxlee 2022-09-09 10:59:55 -05:00
  • d25e6f98e9
    Update prepare_lexicon.py shcxlee 2022-09-09 10:59:35 -05:00
  • fe20d3d222
    Update prepare_transcripts.py shcxlee 2022-09-09 10:57:40 -05:00
  • 27b57bd672
    Update prepare_lexicon.py shcxlee 2022-09-09 10:56:54 -05:00
  • 11680a54fd Modified prepare_transcripts.py and preprare_lexicon.py Seunghyun Lee 2022-09-08 16:35:50 -05:00
  • ac80318a60 Merge branch 'master' into latency pkufool 2022-09-08 11:37:03 +08:00
  • af5a7a400f Adding symlink AmirHussein96 2022-09-08 06:32:17 +03:00
  • 9783de578c Penalty after warmup pkufool 2022-09-08 11:29:57 +08:00
  • 042d4f43e4 .nfs removed AmirHussein96 2022-09-08 04:35:22 +03:00
  • d21dbd5416
    Update asr_datamodule.py Amir Hussein 2022-09-07 21:19:19 -04:00
  • ec365c8d5f
    Update prepare_lang_bpe.py Amir Hussein 2022-09-07 21:10:58 -04:00
  • fc45d7d060
    Update RESULTS.md Amir Hussein 2022-09-07 21:02:16 -04:00
  • 53b0b0c29c
    Update README.md Amir Hussein 2022-09-07 21:00:51 -04:00
  • a89ae13293 stateless transducer MGB-2 AmirHussein96 2022-09-07 19:35:51 +03:00
  • cb840d66c6 . AmirHussein96 2022-09-07 18:14:26 +03:00
  • 7c798de528 Merge branch 'mgb2' of https://github.com/AmirHussein96/icefall into mgb2 AmirHussein96 2022-09-07 18:13:35 +03:00
  • 5f9ef7b04e
    Merge branch 'k2-fsa:master' into mgb2 Amir Hussein 2022-09-07 11:12:29 -04:00
  • 64d6ec0690 Merge branch 'master' of https://github.com/AmirHussein96/icefall into mgb2 AmirHussein96 2022-09-07 18:11:13 +03:00
  • 73e5ba7baa delete comments yaozengwei 2022-09-07 22:09:27 +08:00
  • c2e43c7014 add cutoff for grad filter yaozengwei 2022-09-07 22:06:51 +08:00
  • 902c3036da refact getting median value yaozengwei 2022-09-07 11:56:18 +08:00
  • 890cd1ab75 fix bugs yaozengwei 2022-09-06 10:23:40 +08:00
  • b18850721d add gradient filter yaozengwei 2022-09-05 22:38:39 +08:00
  • 2cc6137934 init files yaozengwei 2022-09-05 22:18:12 +08:00
  • a5f57a4bf2 add gradient filter module yaozengwei 2022-09-05 18:02:54 +08:00
  • 2d8f1c7b7c init files yaozengwei 2022-09-05 14:40:51 +08:00
  • fb2a866e1f add clip grad from #560 Fangjun Kuang 2022-09-04 14:59:43 +08:00
  • 9dd93252cb Merge remote-tracking branch 'zengwei/lstm_clip_gradient' into lstm-giga-libri-clip-grad Fangjun Kuang 2022-09-04 14:30:00 +08:00
  • b8c6532a1a rename Fangjun Kuang 2022-09-04 14:29:18 +08:00
  • 7d5f491e3a fix the gradient clipper yaozengwei 2022-09-04 00:00:41 +08:00
  • 3d931e3386 add missing file Fangjun Kuang 2022-09-03 18:50:05 +08:00
  • 20ef6bb337 fix style issues Fangjun Kuang 2022-09-03 18:46:22 +08:00
  • d1650f6590 Add results Fangjun Kuang 2022-09-03 18:37:02 +08:00
  • fb90ada9e8
    remove trailing white space rickychanhoyin 2022-08-31 13:45:35 +08:00
  • 862e817442
    Very minor change in alimeeting recipe rickychanhoyin 2022-08-31 13:12:28 +08:00
  • ebb1dea786
    Merge branch 'k2-fsa:master' into master rickychanhoyin 2022-08-31 13:00:17 +08:00
  • 0ca1ecde58 fix typo yaozengwei 2022-08-31 12:07:09 +08:00
  • c62851234c clip rnn gradients at each chunk yaozengwei 2022-08-31 11:44:31 +08:00
  • e3128cbccb Add LSTM for the multi-dataset setup. Fangjun Kuang 2022-08-29 19:05:41 +08:00
  • b1cc8e2ec3 Resolve merge Teo 2022-08-29 19:53:06 +09:00
  • 98aaa8bb16 Dockerfile Teo 2022-08-29 19:49:56 +09:00
  • a4dada4ea7
    Merge branch 'k2-fsa:master' into master Teo Wen Shen 2022-08-29 18:48:30 +08:00
  • 0a1ae23944 init files yaozengwei 2022-08-29 15:23:52 +08:00
  • 077719c9ab
    Merge branch 'k2-fsa:master' into master Zengwei Yao 2022-08-29 15:18:11 +08:00
  • e18fa78c3a
    Check that read_manifests_if_cached returns a non-empty dict. (#555) Fangjun Kuang 2022-08-28 11:50:11 +08:00
  • 6e80ca9056 Check that read_manifests_if_cached returns a non-empty dict. Fangjun Kuang 2022-08-28 11:35:35 +08:00
  • d68b8e9120
    Disable CUDA_LAUNCH_BLOCKING in wenetspeech recipes. (#554) Fangjun Kuang 2022-08-28 11:17:38 +08:00
  • 85d12259db minor fixes Fangjun Kuang 2022-08-28 11:16:22 +08:00
  • b5e577fd4e Disable CUDA_LAUNCH_BLOCKING in wenetspeech recipes. Fangjun Kuang 2022-08-27 19:25:05 +08:00
  • 235eb0746f
    fix scaling converter test for decoder(predictor). (#553) kobenaxie 2022-08-27 17:26:21 +08:00
  • 2636a3dd58
    minor changes for correct path names && import module text2segments.py (#552) rickychanhoyin 2022-08-27 17:23:45 +08:00
  • 193f8af38d
    fix scaling converter test for decoder(predictor). kobenaxie 2022-08-27 17:14:19 +08:00
  • 215c4bde15 minor changes for correct path names && import module text2segments.py rickychanhoyin 2022-08-27 16:39:05 +08:00
  • 4237eeabbe
    Merge branch 'k2-fsa:master' into master rickychanhoyin 2022-08-27 15:53:59 +08:00
  • 8bd92c6f18
    Merge branch 'k2-fsa:master' into lightweight Tiance Wang 2022-08-26 14:10:35 +08:00
  • ebcc8e4e1d
    modified beam search with decoder output cached. kobenaxie 2022-08-25 17:53:50 +08:00
  • 26eee02353 change states tensor to batch first for triton Yuekai Zhang 2022-08-25 08:07:46 +00:00
  • c8742ff5e9 warp the list of tensors input Yuekai Zhang 2022-08-25 07:50:58 +00:00
  • 1e31fbcd7d
    Add clamping operation in Eve optimizer for all scalar weights to avoid (#550) marcoyang1998 2022-08-25 12:12:50 +08:00
  • 3ad7dab79c Add clamping operation in Eve optimizer for all scalar weights to avoid non stable training in some scenarios. The clamping range is set to (-10,2). Note that this change may cause unexpected effect if you resume training from a model that is trained without clamping. marcoyang 2022-08-25 11:38:04 +08:00
  • bf6c5601c7 Support finetuning hubert transducer in icefall marcoyang 2022-08-25 11:33:43 +08:00
  • 0967cf5b38
    fixed no cut_id error in decode_dataset (#549) Duo Ma 2022-08-25 10:54:21 +08:00
  • bd2b455f12 fixed code style shanguanma 2022-08-25 10:47:48 +08:00
  • c84a0313f4 Merge branch 'master' of https://github.com/shanguanma/icefall shanguanma 2022-08-25 10:01:33 +08:00
  • dfc45581b3 fixed more than one "#" shanguanma 2022-08-25 10:00:19 +08:00
  • 21f3f72326
    Merge branch 'k2-fsa:master' into master Duo Ma 2022-08-25 09:56:00 +08:00
  • 92b2ded2a9 fixed no cut_id error in decode_dataset shanguanma 2022-08-25 09:52:49 +08:00