Commit Graph

  • 9f2229edb5 Merge making hidden_dim an arg Daniel Povey 2022-08-02 08:58:00 +08:00
  • a45f820e25 Merge making hidden_dim an arg Daniel Povey 2022-08-02 08:56:36 +08:00
  • 6714f85cc4 Merge making hidden_dim an arg Daniel Povey 2022-08-02 08:55:27 +08:00
  • 804f264ffd Merge hidden_dim providing it as arg Daniel Povey 2022-08-02 06:33:31 +08:00
  • 6e43a2b69d Remove ncnn as it is not fully tested yet. Fangjun Kuang 2022-08-01 19:51:38 +08:00
  • 0d5edbc3f2 Add more doc. Fangjun Kuang 2022-08-01 19:51:23 +08:00
  • ee311247ea Decrease debugging freq Daniel Povey 2022-08-01 03:55:18 +08:00
  • 4c5d49c448 Some numerical improvements, and a fix to calculation of mean_eig in _apply_min_max_with_metric(), to average over blocks too. Daniel Povey 2022-08-01 03:51:39 +08:00
  • 5313ce00d6 simplify relative position encoding yaozengwei 2022-07-31 20:31:02 +08:00
  • b338471917 copy pruned4, with streaming mode yaozengwei 2022-07-31 20:24:34 +08:00
  • e2cc09a8c6 Fix issue with max_eig formula; restore cov_min[1]=0.0025. Daniel Povey 2022-07-31 18:29:44 +08:00
  • b8abf38aca Merge remote-tracking branch 'k2-fsa/master' into pos_enc yaozengwei 2022-07-31 18:14:32 +08:00
  • 3590c2fc42 Set cov_min[1] to 0 to stop an invertibility problem Daniel Povey 2022-07-31 18:05:36 +08:00
  • 7231c610e8 Restore min_cov applied with G. Daniel Povey 2022-07-31 02:22:07 -07:00
  • d84a2e22e3 Applying max to G with noinv method with metric. Daniel Povey 2022-07-31 02:10:27 -07:00
  • 2042c9862c Merge branch 'pradam_exp1m4_nophase1_noinv' into pradam_exp1m4_nophase1_rework_noinv Daniel Povey 2022-07-31 01:32:36 -07:00
  • 90fa8a63eb Use different approach for applying max eig, with matmul, no inverse. Daniel Povey 2022-07-31 01:32:11 -07:00
  • ed1a147ef1 Implement no-inverse max-cov Daniel Povey 2022-07-31 00:08:02 -07:00
  • 0666789cb8 Small numerical improvements; config change of eps and G_diag changed 1.01 to 1.005; decrease an eps from 1e-10 to 1e-20 Daniel Povey 2022-07-30 21:48:54 -07:00
  • cb67540cdc this version not working great Daniel Povey 2022-07-30 21:14:03 -07:00
  • 790e8c4ba9 Changes that should not really affect the results, just cleanup. Daniel Povey 2022-07-30 19:20:36 -07:00
  • 5184ac570d Removing phase1, adding regular smoothing with the mean. Daniel Povey 2022-07-30 19:15:51 -07:00
  • e1182da6ac Restoring min_abs and max_abs defaults for the linear_pos proj. Daniel Povey 2022-07-31 05:07:50 +08:00
  • eb549bf5d0 Fix typos. Fangjun Kuang 2022-07-30 22:19:37 +08:00
  • f572e149a9 Decode with exported models. Fangjun Kuang 2022-07-30 21:17:31 +08:00
  • 105d49d31b Big simplification to update rule Daniel Povey 2022-07-30 00:21:12 -07:00
  • 71ea196370 Update CI to test ONNX models. Fangjun Kuang 2022-07-30 11:50:39 +08:00
  • c70df281c6 Revert merging three onnx models into a single one. Fangjun Kuang 2022-07-30 11:33:22 +08:00
  • a80a8abf0c Merge branch 'pradam_exp1l3' into pradam_exp1m3 Daniel Povey 2022-07-30 09:07:05 +08:00
  • 4d5323d6a3 Change cov_min to intermediate value, not seeing clear diffs from changing it Daniel Povey 2022-07-30 09:06:23 +08:00
  • c0e652ea87 Fix code, was smoothing in wrong basis. Daniel Povey 2022-07-30 09:01:11 +08:00
  • c5a037b8bc Merge branch 'pradam_exp1l3' into pradam_exp1m3 Daniel Povey 2022-07-30 08:21:28 +08:00
  • 17bc002e6e Refactoring that does not affect results. Daniel Povey 2022-07-30 07:45:29 +08:00
  • 3110138ab5 Smooth grad_cov with eps; add a 4th stage of smoothing, this time on Z_inv. Daniel Povey 2022-07-30 07:30:32 +08:00
  • 3ebb52aa9b Combine encoder/decoder/joiner into a single file. Fangjun Kuang 2022-07-29 21:14:20 +08:00
  • 2f75236c05
    Support dynamic chunk streaming training in pruned_transcuder_stateless5 (#454) Wei Kang 2022-07-29 16:40:06 +08:00
  • 0229d87b12 fix streaming decode; add results pkufool 2022-07-29 16:02:40 +08:00
  • ca28f46f75 Merge branch 'pradam_exp1l2' into pradam_exp1m2 Daniel Povey 2022-07-29 15:16:10 +08:00
  • 3ad042444e More changes to reduce numerical roundoff for dims with zero grad and params. Daniel Povey 2022-07-29 14:38:50 +08:00
  • c7920bf303 Merge branch 'master' into streaming5 pkufool 2022-07-29 14:38:23 +08:00
  • 1b478d3ac3
    Add other decoding methods (nbest, nbest oracle, nbest LG) for wenetspeech pruned rnnt2 (#482) Mingshuang Luo 2022-07-29 12:03:08 +08:00
  • 4e2a49d102
    add nbest-scale for flie name Mingshuang Luo 2022-07-29 11:53:19 +08:00
  • a55f8c9c14 Modify scaling.py to prevent constant values Daniel Povey 2022-07-29 09:34:13 +08:00
  • 34b4356bad
    correction for get rank id. (#507) Lucky Wong 2022-07-29 11:28:52 +08:00
  • 1e3a0e09c6
    Update README.md Mingshuang Luo 2022-07-29 11:20:11 +08:00
  • d64fb92f16
    set ngram-lm-scale=0.35 as default Mingshuang Luo 2022-07-29 11:17:31 +08:00
  • ec69967584
    Set overwrite=True when extracting features in batches. (#487) Fangjun Kuang 2022-07-29 11:17:19 +08:00
  • 9d7af4be20 Modify scaling.py to prevent constant values Daniel Povey 2022-07-29 09:34:13 +08:00
  • 3c1fddaf48 Rework computation to reduce numerical roundoff Daniel Povey 2022-07-29 06:22:17 +08:00
  • ee8397e961 Minor fixes pkufool 2022-07-28 22:32:51 +08:00
  • 49187bc29d
    correction for get rank id. Lucky Wong 2022-07-28 22:15:28 +08:00
  • 0cffe78485
    Merge branch 'k2-fsa:master' into master Lucky Wong 2022-07-28 22:13:15 +08:00
  • 83e2b30a22 adding black reformatted files s-mousmita 2022-07-28 07:58:20 -04:00
  • 389f9c77e5
    correction for prepare.sh (#506) Mingshuang Luo 2022-07-28 17:01:46 +08:00
  • 196a7c404f correction for prepare.sh luomingshuang 2022-07-28 16:46:58 +08:00
  • 3c9e7f733b
    [debug] raise remind when git-lfs not available (#504) boji123 2022-07-28 16:17:49 +08:00
  • 43374aae93 modify comment boji123 2022-07-28 16:09:03 +08:00
  • 49aaaf8021 Minor fixes. Fangjun Kuang 2022-07-28 15:52:51 +08:00
  • 1ee36f28e5 [debug] raise remind when git-lfs not available boji123 2022-07-28 15:48:17 +08:00
  • 633cbd551a Increase lr_update_period from 200,4000 to 400, 5000 Daniel Povey 2022-07-28 14:45:45 +08:00
  • f26b62ac00
    [WIP] Pruned-transducer-stateless5-for-WenetSpeech (offline and streaming) (#447) Mingshuang Luo 2022-07-28 12:54:27 +08:00
  • eb578f9fb8 add streaming results on README.md luomingshuang 2022-07-28 12:30:57 +08:00
  • 4b07301e15 add some decode commands for usage luomingshuang 2022-07-28 11:58:11 +08:00
  • 44ec59fc1f do some changes for export.py luomingshuang 2022-07-28 10:22:26 +08:00
  • 0d038a6ea4 Remove debugging statement Daniel Povey 2022-07-28 09:25:10 +08:00
  • 8654a7385d Add denom_rel_eps, and set it to 1e-05 Daniel Povey 2022-07-28 09:10:20 +08:00
  • dc565f729b Take into account various outcomes from parameter tuning Daniel Povey 2022-07-28 09:06:59 +08:00
  • 8c98599ded WIP: Support exporting to ONNX format Fangjun Kuang 2022-07-27 22:58:27 +08:00
  • 23be42a8c9 do some changes luomingshuang 2022-07-27 19:21:01 +08:00
  • d2581cfcb4 change for style_check.yml luomingshuang 2022-07-27 17:59:55 +08:00
  • ffb25d12dc add README.md and RESULTS.md luomingshuang 2022-07-27 16:41:23 +08:00
  • daa55d5a3c Patches to make decoding work correctly at utt start, for greedy_search Daniel Povey 2022-07-27 09:32:51 +08:00
  • 9cf447646d modify the doc of RelPositionMultiheadAttention yaozengwei 2022-07-26 22:31:06 +08:00
  • 67e3607863 Updating musan part to make compatible wiht recent lhotse nomenclature s-mousmita 2022-07-26 10:30:34 -04:00
  • 1f843e18f2 Merge remote-tracking branch 'dan/master' into quantization Fangjun Kuang 2022-07-26 16:50:04 +08:00
  • cc44466854 Add modified_beam_search and fast_beam_search. Fangjun Kuang 2022-07-26 16:49:34 +08:00
  • a5a7514ef0 change for streaming_beam_search.py luomingshuang 2022-07-26 16:47:38 +08:00
  • 385645d533
    Fix get_transducer_model() for aishell. (#497) Fangjun Kuang 2022-07-26 15:42:21 +08:00
  • 5938dd46e5 Fix get_transducer_model() for aishell. Fangjun Kuang 2022-07-26 15:40:23 +08:00
  • 93f4eb1164 add ngram-lm-scale=0.35 results luomingshuang 2022-07-26 14:56:03 +08:00
  • e25ca74955 Use a measure of correlation for eigs that can be negative. Daniel Povey 2022-07-26 13:40:57 +08:00
  • c03d7f0079 Merge branch 'master' into latency pkufool 2022-07-26 13:37:54 +08:00
  • 718086460e training pruned_transducer_stateless4 with delay-penalty pkufool 2022-07-26 13:31:38 +08:00
  • b9696878b4 Update diagnostics stats Daniel Povey 2022-07-26 12:30:07 +08:00
  • d3fc4b031e
    Support using aidatatang_200zh optionally in aishell training (#495) Fangjun Kuang 2022-07-26 11:25:01 +08:00
  • 06377c0adf minor fixes Fangjun Kuang 2022-07-26 11:21:21 +08:00
  • bb37894fa9 minor fixes Fangjun Kuang 2022-07-26 11:00:03 +08:00
  • b34eafa500 Use aidatatang_200zh optionally in aishell training. Fangjun Kuang 2022-07-26 10:55:56 +08:00
  • 4612b03947
    Fix using G before assignment in pruned_transducer_stateless/decode.py (#494) Fangjun Kuang 2022-07-26 10:37:02 +08:00
  • ca9be09cc1 Use mode.eval() in decoding. Fangjun Kuang 2022-07-25 21:19:33 +08:00
  • c1fc004df5 minor fixes Fangjun Kuang 2022-07-25 21:10:27 +08:00
  • 173a3c48c9 Fix using G before assignment in pruned_transducer_stateless/decode.py Fangjun Kuang 2022-07-25 20:45:52 +08:00
  • 90dc5772ec Support decoding from a torchscript model. Fangjun Kuang 2022-07-25 20:34:13 +08:00
  • 6871c96ffa add using proj_size yaozengwei 2022-07-25 19:04:08 +08:00
  • 34524acf44 Removing some run time comment s-mousmita 2022-07-25 07:03:54 -04:00
  • 2777c0b0b3 Support quantization Fangjun Kuang 2022-07-25 18:14:21 +08:00
  • b1d0956855
    Add modified_beam_search for streaming decode (#489) Wei Kang 2022-07-25 16:53:23 +08:00
  • 8203d10be7
    Add stats about duration and padding proportion (#485) Zengwei Yao 2022-07-25 16:40:43 +08:00
  • 691deaf730 minor change yaozengwei 2022-07-25 16:31:05 +08:00
  • 4bf839227c add modified-beam-search for streaming decoding luomingshuang 2022-07-25 16:11:59 +08:00