660facf | freewym | 20 July 2023, 23:05:26 UTC | allows dictionary files w/o the counts column; rename task's --max-num-expansions-per-step to --transducer-max-num-expansions-per-step (same as generation's) and its default is 20; prints out word counts after WER evaluation; fixes decoding log write out | 24 July 2023, 00:01:56 UTC |
e0e61e2 | freewym | 14 July 2023, 01:36:27 UTC | add options to be able to share learnable positional embeddings across layers or attention heads | 14 July 2023, 01:36:27 UTC |
929ad76 | Karel Vesely | 02 December 2022, 14:26:57 UTC | Bugfix in transformer model creation. Making the signatures of methods compatible: `SpeechTransformerModelBase.build_encoder(...)` `SpeechTransformerModel.build_encoder(...)` (otherwise, the model build crashing) | 08 December 2022, 09:00:45 UTC |
09c7a91 | freewym | 10 November 2022, 01:41:23 UTC | makes BatchNorm configurable in preencoder; adds the bias term to joiner's final fc_out and uses its default initializer | 08 December 2022, 09:00:45 UTC |
f823f17 | freewym | 30 October 2022, 08:48:26 UTC | adds an option to include EOS in transducer model training; fixes a bug regarding attn_mask; rewrites masked_copy_cached_state() to make it clearer and more general; removes clone_cached_state(); code adaptation/changes according to the commits on Nov 2, 2022 | 08 December 2022, 09:00:45 UTC |
f45c936 | freewym | 28 October 2022, 22:00:19 UTC | transducer decoding fix; cosmatic changes | 08 December 2022, 09:00:44 UTC |
3b60635 | freewym | 18 October 2022, 03:12:14 UTC | code adaptation/changes according to the commits on Oct 17, 2022 | 08 December 2022, 09:00:44 UTC |
54df33a | freewym | 10 October 2022, 05:47:16 UTC | adds chunk streaming mask support for Transformer | 08 December 2022, 09:00:44 UTC |
3c21b82 | freewym | 29 September 2022, 00:47:35 UTC | handle installation of packaging | 08 December 2022, 09:00:44 UTC |
7fa5d35 | freewym | 21 September 2022, 22:18:10 UTC | code adaptation/changes according to the commits on Sep 8, 2022 | 08 December 2022, 09:00:44 UTC |
4834687 | Yiming Wang | 21 September 2022, 21:20:14 UTC | CTC model is supported; adjusts num_workers and required_batch_size_m… (#76) * CTC model is supported; adjusts num_workers and required_batch_size_multiple; adjusts num updates accordingly; cosmetic changes | 08 December 2022, 09:00:43 UTC |
4bf0874 | freewym | 12 August 2022, 20:16:58 UTC | cosmetic improvement for specaugment config and transducer loss; add extra_symbols_to_ignore when printing out sample predictions | 08 December 2022, 09:00:43 UTC |
3c5be1b | freewym | 01 August 2022, 23:01:51 UTC | Revert "code adaptation/changes according to the commits on June 15, 2022" This reverts commit 63d3458de0e6846e95e992e738f2a823d8a9ffd8. | 08 December 2022, 09:00:43 UTC |
45b30ae | freewym | 16 June 2022, 07:58:26 UTC | code adaptation/changes according to the commits on June 15, 2022 | 08 December 2022, 09:00:43 UTC |
844c858 | freewym | 13 May 2022, 23:21:50 UTC | add the "learnable" property for relative positional embedding and enquiry that in MultiheadAttention | 08 December 2022, 09:00:43 UTC |
5a2b90e | freewym | 06 May 2022, 07:48:20 UTC | code adaptation/changes according to the commits on May 4, 2022 | 08 December 2022, 09:00:43 UTC |
77621cc | freewym | 25 April 2022, 09:11:28 UTC | code adaptation/changes according to the commits on Apr 20, 2022 | 08 December 2022, 09:00:43 UTC |
99b272e | freewym | 27 March 2022, 09:05:19 UTC | code adaptation/changes according to the commits on Mar 24, 2022 | 08 December 2022, 09:00:43 UTC |
16900d9 | freewym | 14 March 2022, 05:12:46 UTC | adaptive specaugment | 08 December 2022, 09:00:43 UTC |
312ac68 | freewym | 13 March 2022, 07:37:08 UTC | use feature_transforms' API for applying SpecAugment; get rid of hard-coding the train_subset name in tasks | 08 December 2022, 09:00:42 UTC |
4204aac | Yiming Wang | 06 March 2022, 07:02:37 UTC | conformer (#73) | 08 December 2022, 09:00:42 UTC |
bd2961f | freewym | 30 January 2022, 03:56:30 UTC | code adaptation/changes according to the commits on Jan 20, 2022 | 08 December 2022, 09:00:41 UTC |
13c2817 | freewym | 15 January 2022, 00:47:59 UTC | update relative positional embedding implementations; tune a bit of transducer models | 08 December 2022, 09:00:41 UTC |
79d5a2b | freewym | 04 January 2022, 09:53:16 UTC | batch data based on both src and tgt lengths for transducer models | 08 December 2022, 09:00:40 UTC |
c941705 | Yiming Wang | 31 December 2021, 22:47:45 UTC | transformer transducer code check-in (#72) Includes: * a transformer-transducer model * transducer loss * Noam lr scheduler * greedy decoding and beam search decoding with LM shallow fusion * a LibriSpeech recipe | 08 December 2022, 09:00:40 UTC |
6f4558a | freewym | 22 December 2021, 09:23:18 UTC | in-place masked_copy_incremental_state; fix import module issue by dummy importing espresso in sequence_generator.py | 08 December 2022, 09:00:40 UTC |
f9220e9 | freewym | 20 December 2021, 09:54:37 UTC | remove an unnecessary check for LM in sequence_generator.py; other cosmetic changes | 08 December 2022, 09:00:39 UTC |
5c4c288 | freewym | 10 December 2021, 03:52:23 UTC | suppress deprecation warning for division with rounding | 08 December 2022, 09:00:39 UTC |
2ead96c | freewym | 03 December 2021, 01:49:18 UTC | reformat for the previous commit | 08 December 2022, 09:00:39 UTC |
c3f66c2 | Timo Lohrenz | 03 December 2021, 01:41:29 UTC | added relaxed attention to encoder-decoder attention (arxiv.org/abs/2107.01275) (#71) * added relaxed attention to encoder-decoder attention (arxiv.org/abs/2107.01275) * Adressed change requests for PR * override build_encoder_attn for relaxed attention PR Co-authored-by: Timo Lohrenz <lohrenz@ifn.ing.tu-bs.de> | 08 December 2022, 09:00:39 UTC |
2cded21 | freewym | 30 November 2021, 02:28:25 UTC | code adaptation/changes according to the commits on Nov 29, 2021 (Add linting with black) | 08 December 2022, 09:00:38 UTC |
c5518d3 | freewym | 13 October 2021, 21:41:32 UTC | code adaptation/changes according to the commits on Oct 13, 2021 | 08 December 2022, 09:00:38 UTC |
de30590 | freewym | 09 September 2021, 19:21:11 UTC | code adaptation/changes according to the commits on Sep 9, 2021 | 08 December 2022, 09:00:38 UTC |
9c3c820 | freewym | 07 August 2021, 09:47:33 UTC | refactor transformer 2 (according to the commits on July 16-28, 2021) | 08 December 2022, 09:00:38 UTC |
7c5037f | freewym | 04 August 2021, 01:46:33 UTC | refactor transformer 1 (according to the commits on July 16, 2021) | 08 December 2022, 09:00:38 UTC |
9fb004a | freewym | 25 June 2021, 02:16:17 UTC | code adaptation/changes according to the commits on June 24-July 1, 2021 | 08 December 2022, 09:00:38 UTC |
6e73519 | Yiming Wang | 14 June 2021, 08:41:44 UTC | Relative Positional Embedding (fixed or learnable) (#68) * Relative Positional Embedding (fixed or learnable) * support relative positional embeddings in transformer models * bug fixes; fixes command line args passing for SpeechTransformerModel * f * ff | 08 December 2022, 09:00:38 UTC |
c14194a | freewym | 11 June 2021, 06:25:32 UTC | code adaptation/changes according to the commits on June 11, 2021 | 08 December 2022, 09:00:38 UTC |
e5eff0e | freewym | 04 June 2021, 07:46:05 UTC | code adaptation/changes according to the commits on June 3, 2021; empty GPU memory cache of the WSJ recipe more frequently to avoid occasional CUDNN error; fixes the command line args parsing problem for SpeechTransformerModel | 08 December 2022, 09:00:37 UTC |
b3de88b | freewym | 30 May 2021, 01:55:35 UTC | fix a comment about the shape of encoder_padding_mask in speech_lstm; fix a bug in the task speech_recognition_hybrid about converting to an iterator; fix an import bug in feat_text_dataset | 08 December 2022, 09:00:37 UTC |
81d86a1 | freewym | 21 May 2021, 06:46:59 UTC | code adaptation/changes according to the commits on May 21, 2021 | 08 December 2022, 09:00:37 UTC |
c5fcfa1 | freewym | 20 May 2021, 03:31:05 UTC | decapitalize bool values in yaml configs | 08 December 2022, 09:00:36 UTC |
026faf1 | freewym | 01 May 2021, 08:33:54 UTC | add an option to allow multi-layer LSTM defined as a single module (faster and may fix a possible cuDNN error when defined as a ModuleList); add missing global cmvn to the decoding stage | 08 December 2022, 09:00:36 UTC |
e4213c7 | Yiming Wang | 25 April 2021, 04:13:32 UTC | hydra migration for librispeech torchaudio recipes (#65) | 08 December 2022, 09:00:36 UTC |
76106fa | freewym | 23 April 2021, 19:45:06 UTC | code adaptation/changes according to the commits on Apr 23, 2021 | 08 December 2022, 09:00:36 UTC |
0184238 | freewym | 21 April 2021, 05:56:47 UTC | more informative error msgs when corpus dirs are not specified | 08 December 2022, 09:00:36 UTC |
dcae0c8 | freewym | 20 April 2021, 21:15:53 UTC | move some task-specifc code in speech_train.py to task functions; move two remaining ones to fairseq's train.py and remove speech_train.py | 08 December 2022, 09:00:36 UTC |
096cddc | Yiming Wang | 20 April 2021, 01:34:22 UTC | on-the-fly feature extraction with torchaudio + LibriSpeech recipe without dependency on Kaldi (#61) | 08 December 2022, 09:00:36 UTC |
6ff8231 | freewym | 15 April 2021, 09:36:38 UTC | fix misleading error msgs on non-CLSP machines; normalize LM training by update_freq; other style changes | 08 December 2022, 09:00:36 UTC |
236c0d0 | freewym | 03 April 2021, 19:09:48 UTC | several bug fixes for hybrid systems (thanks to liuhongji2000) | 08 December 2022, 09:00:36 UTC |
0939b1d | freewym | 31 March 2021, 19:49:13 UTC | support bos in class AsrDictionary for CTC loss, etc | 08 December 2022, 09:00:36 UTC |
1edb65c | freewym | 30 March 2021, 05:49:02 UTC | code adaptation/changes according to the commits on Mar 29-30, 2021 | 08 December 2022, 09:00:35 UTC |
1b603c6 | freewym | 26 March 2021, 22:33:45 UTC | on-the-fly tokenization | 08 December 2022, 09:00:35 UTC |
6102cde | freewym | 23 March 2021, 04:04:48 UTC | code adaptation/changes according to the commits on Mar 22-23, 2021 | 08 December 2022, 09:00:35 UTC |
0916351 | freewym | 22 March 2021, 18:32:26 UTC | move convolutions from lstm to a separate file | 08 December 2022, 09:00:35 UTC |
9dca701 | freewym | 13 March 2021, 08:20:56 UTC | fixes a bug in FeatScpCachedDataset; code adaptation/changes according to the commits on Mar 20, 2021 | 08 December 2022, 09:00:35 UTC |
86cf6de | freewym | 05 March 2021, 03:38:54 UTC | code adaptation/changes according to the commits on Mar 4-12, 2021 | 08 December 2022, 09:00:35 UTC |
a291aeb | Yiming Wang | 03 March 2021, 11:01:55 UTC | tokenization within datasets (need to re-run json prep stage right before training stage to be compatible with this change) (#58) | 08 December 2022, 09:00:35 UTC |
3c8712a | freewym | 09 February 2021, 10:02:21 UTC | code adaptation/changes according to the commits on Feb 8-Mar 2, 2021 | 08 December 2022, 09:00:35 UTC |
537b3b2 | freewym | 03 February 2021, 23:05:33 UTC | fix labeled smoothed cross entropy computation according to the corresponding changes in fairseq on Feb 3rd, 2021; affects recipes using uniform smoothing | 08 December 2022, 09:00:34 UTC |
a3ae124 | freewym | 21 January 2021, 22:07:29 UTC | code adaptation/changes according to the commits on Jan 20-28, 2021 | 08 December 2022, 09:00:34 UTC |
a9558b3 | freewym | 28 December 2020, 21:29:20 UTC | fix typos in argument names; code adaptation/changes according to the commits on Dec 28, 2020 - Jan 5, 2021 | 08 December 2022, 09:00:34 UTC |
a8b1c99 | freewym | 23 December 2020, 08:53:50 UTC | code adaptation/changes according to the commits on Dec 22, 2020 | 08 December 2022, 09:00:34 UTC |
907c9bc | freewym | 04 December 2020, 20:52:07 UTC | fix length tensor device issue in lf_mmi loss; code adaptation/changes according to the commits on Dec 3-12, 2020 | 08 December 2022, 09:00:34 UTC |
f351215 | freewym | 18 November 2020, 21:57:26 UTC | code adaptation/changes according to the commits on Nov 16-20, 2020; fix a bug in Multi-level LM when getting cached states | 08 December 2022, 09:00:33 UTC |
83390bf | freewym | 14 November 2020, 22:33:49 UTC | fix an error when more than one external LMs are used for shallow fusion | 08 December 2022, 09:00:33 UTC |
976dfa8 | freewym | 11 November 2020, 21:38:40 UTC | code adaptation/changes according to the commits on Nov 11, 2020; obtain feat_dim in setup_task() instead | 08 December 2022, 09:00:33 UTC |
4734e37 | freewym | 05 November 2020, 05:52:02 UTC | code adaptation/changes according to the commits on Nov 4-9, 2020 | 08 December 2022, 09:00:33 UTC |
8028172 | freewym | 26 October 2020, 02:27:48 UTC | code adaptation/changes according to the commits on Oct 18-Nov 3, 2020 (lots of changes, mostly for adapting to hydra configs and code formatting) | 08 December 2022, 09:00:33 UTC |
91973ad | freewym | 11 October 2020, 23:48:30 UTC | code adaptation/changes according to the commits on Oct 2-15, 2020 | 08 December 2022, 09:00:32 UTC |
63bdcdf | freewym | 01 October 2020, 20:19:56 UTC | code adaptation/changes according to the commits on Oct 1, 2020 | 08 December 2022, 09:00:32 UTC |
3987af6 | freewym | 18 September 2020, 08:12:09 UTC | code adaptation/changes according to the commits on Sep 17-26, 2020 | 08 December 2022, 09:00:32 UTC |
f5bd61f | freewym | 09 September 2020, 16:30:45 UTC | code adaptation/changes according to the commits on Sep 9-11, 2020 | 08 December 2022, 09:00:31 UTC |
094ecc7 | freewym | 31 August 2020, 20:02:27 UTC | code adaptation/changes according to the commits on Aug 31, 2020 | 08 December 2022, 09:00:31 UTC |
38b199c | freewym | 21 August 2020, 04:20:39 UTC | code adaptation/changes according to the commits on Aug 20-24, 2020 | 08 December 2022, 09:00:31 UTC |
eb6507e | freewym | 10 August 2020, 19:07:42 UTC | code adaptation/changes according to the commits on Aug 10-18, 2020 | 08 December 2022, 09:00:31 UTC |
ef05c79 | Yiming Wang | 10 August 2020, 00:44:46 UTC | updates for new PyChain (#37) * add support for output l2 regularization and xent regularization; add a bichar WSJ recipe; add missing soft links to kaldi files * move ChainLossFunction here from PyChain | 08 December 2022, 09:00:31 UTC |
509be47 | freewym | 30 July 2020, 23:49:54 UTC | reorder the elements of the returned tuple of TdnnModel.forward(); export KALDI_ROOT to adapt to the recent changes in kaldi_io; code adaptation/changes according to the commits on Aug 3-4, 2020 | 08 December 2022, 09:00:31 UTC |
e916ef5 | freewym | 28 July 2020, 22:42:39 UTC | fix reorder_encoder_out in SpeechChunkTransformerEncoder; code adaptation/changes according to the commits on Jul 28, 2020 | 08 December 2022, 09:00:31 UTC |
271bd2a | freewym | 21 July 2020, 03:16:27 UTC | code adaptation/changes according to the commits on Jul 20-25, 2020 | 08 December 2022, 09:00:31 UTC |
33e5bd9 | freewym | 17 July 2020, 01:56:03 UTC | code adaptation/changes according to the commits on Jul 16, 2020 | 08 December 2022, 09:00:31 UTC |
7b491aa | freewym | 15 July 2020, 20:38:39 UTC | code adaptation/changes according to the commits on Jul 14, 2020 | 08 December 2022, 09:00:31 UTC |
fdc42c1 | freewym | 08 July 2020, 01:25:29 UTC | ignore flake8's FileNotFoundError for soft links to kaldi files; code adaptation/changes according to the commits on Jul 8, 2020 | 08 December 2022, 09:00:31 UTC |
591e19d | Yiming Wang | 02 July 2020, 17:55:20 UTC | Update Transformer models (#31) * update transformer * initial recipe * fix transformer * add encoder positional embeddings * add more recipes | 08 December 2022, 09:00:30 UTC |
35b0e71 | freewym | 25 June 2020, 23:54:03 UTC | code adaptation/changes according to the commits on Jun 24-25, 2020; fix validation loss in LSTM models | 08 December 2022, 09:00:30 UTC |
c09a878 | freewym | 19 June 2020, 05:34:20 UTC | code adaptation/changes according to the commits on Jun 18-23, 2020 | 08 December 2022, 09:00:30 UTC |
08ddb46 | freewym | 17 June 2020, 00:11:09 UTC | remove useless max_{source,target}_positions arguments | 08 December 2022, 09:00:30 UTC |
f7359ad | freewym | 27 May 2020, 04:39:10 UTC | fix lf-mmi loss; code adaptation/changes according to the commits on May 27 | 08 December 2022, 09:00:30 UTC |
8af6d18 | freewym | 19 May 2020, 19:00:51 UTC | code adaptation/changes according to the commits on May 18 | 08 December 2022, 09:00:29 UTC |
95a109d | freewym | 10 May 2020, 19:24:17 UTC | code adaptation/changes according to the commits on May 10 | 08 December 2022, 09:00:29 UTC |
a7c391e | Yiming Wang | 02 May 2020, 23:53:55 UTC | Hybrid ASR code (E2E LF-MMI and cross-entropy) and WSJ examples (#29) | 08 December 2022, 09:00:29 UTC |
c34a7c2 | freewym | 23 April 2020, 20:46:01 UTC | use EncoderOut for SpeechLSTMEncoder's output; code adaptation/changes according to the commits on Apr 21 | 08 December 2022, 09:00:29 UTC |
c748fde | freewym | 14 April 2020, 02:55:40 UTC | update the qsub script for gpu jobs; code adaptation/changes according to the commits on Apr 16 | 08 December 2022, 09:00:28 UTC |
46dcea9 | freewym | 07 April 2020, 22:20:41 UTC | code adaptation/changes according to the commits on Apr 7 | 08 December 2022, 09:00:28 UTC |
6d0eb55 | freewym | 24 March 2020, 20:01:19 UTC | code adaptation/changes according to the commits on Mar 24-Apr 3, 2020; use data.encoders.{bpe,tokenizer} for wordpiece decode | 08 December 2022, 09:00:28 UTC |
444510f | freewym | 20 March 2020, 04:29:42 UTC | fix specaug indexing | 08 December 2022, 09:00:28 UTC |
6366fff | freewym | 11 March 2020, 16:58:26 UTC | code adaptation/changes according to the commits on Mar 11, 2020; change logs/->log/; rename SpeechDataset->AsrDataset, Scp*Dataset->FeatScp*Dataset, score*.sh->score*_e2e.sh; remove validation on train subset from wsj recipe | 08 December 2022, 09:00:28 UTC |
2084287 | Yiming Wang | 11 March 2020, 06:21:30 UTC | SpecAugment (#21) | 08 December 2022, 09:00:28 UTC |
86346e7 | freewym | 05 March 2020, 08:50:15 UTC | code adaptation/changes according to the commits on Mar 3-10, 2020 | 08 December 2022, 09:00:28 UTC |
01c5fbe | freewym | 27 February 2020, 23:07:19 UTC | code adaptation/changes according to the commits on Feb 27-29, 2020 | 08 December 2022, 09:00:27 UTC |