[WIP] Lhotse/K2 example #45
Commits on Dec 24, 2020
-
-
-
-
-
-
-
-
-
-
word lm related; add unigram/temporal label smoothing and update the wsj
recipe accordingly; code adaptation/changes according to the commits from Apr 12, 2019 to May 21, 2019
-
-
-
-
librispeech recipe fix; code adaptation/changes according to the comm…
…its on June 11, 2019
-
-
-
add coverage to libripspeech recipe; make best metric for ASR configu…
…rable by using --best-checkpoint-metric; code adaptation/changes according to the commits from Jun 26, 2019 to Jul 3, 2019
-
fix swbd recipe; code adaptation/changes according to the commits fro…
…m Jul 17, 2019 to Jul 19, 2019
-
-
improve swbd recipe; comtinue training while lr is no less than --min…
…-lr; code adaptation/changes according to the commits from Jul 24, 2019 to Aug 1, 2019
-
-
add --eos-factor for beam search to alleviate the problem of too shor…
…t transcripts with LM fusion
-
tokenize each sentence such that it ends with <space>; modify look-ah…
…ead LM accordingly; modify the WSJ recipe accordingly
-
switch to pip install sentencepiece; modify Librispeech/SWBD recipes …
…accordingly; code adaptation/changes according to the commits on Aug 21 and 30, 2019
-
-
-
Update README.md; add logo; slightly change LM weight and beam size for
Librispeech; code adaptation/changes according to the commits on Sep 17, 2019
-
compansate for the removal of torch.rand() from distributed_init() re…
…cently introduced in fairseq, to make ASR results reproducible
-
-
a better LM for Librispeech yielding better WERs; code
adaptation/changes according to the commits on Sep 27, 2019
-
set --distributed-port=-1 if ngpus=1; code adaptation/changes accordi…
…ng to the commits on Oct 11, 2019
-
change warmup scheduling for ReduceLROnPlateauV2; code adaptation/cha…
…nges according to the commits on Oct 18, 2019
-
remove warmup code in ReduceLROnPlateauV2 as fariseq just added it; c…
…ode adaptation/changes according to the commits on Oct 23, 2019
-
-
Fixed error when using fp16. Followed example from https://github.com/pytorch/fairseq/blob/master/fairseq/models/lstm.py#L299
-
allows text2vocabulary.py to accept an existing vocabulry with its fi…
…rst column as words
-
-
remove coverage term for beam search decoding as it has been superced…
…ed by eos thresholding
-
fix bugs causing build failure; a bunch of lint changes; rename
TokenDictionary->AsrDictionary, TokenTextDataset->AsrTextDataset
-
-
decouple scheduled sampling rate scheduler; rename all appearances of…
… "dict" variables to "dictionary" as "dict" is a reserved keyword in Python; affect swbd results due to some numerical issue of PyTorch
-
move the code of computing prob mask of temporal label smoothing into…
… a separate function
-
isolate greedy search code from criterions (#19)
* isolate greedy search code from criterions
-
remove the need to pass tgt dataset to criterions by adding a raw tex…
…t field to collated samples (#20)
-
code adaptation/changes according to the commits from Jan 16 to Jan 1…
…7, 2020; move decode log files to decode result dirs
-
move WER validation with greedy decoding code from criterion to task;…
… valid loss is now based on teacher enforcing instead of greedy decoding; rename {,label_smoothed_}cross_entropy_with_wer.py to {,label_smoothed_}cross_entropy_v2.py -
code adaptation/changes according to the commits on Jan 20, 2020; cos…
…metic changes for lookahead LM
-
isolate LSTMLanguageModel from speech_lstm.py and rename it to LSTMLa…
…nguageModelEspresso to avoid naming conflicts with the fairseq\'s LSTMLanguageModel introduced recently
-
add options to accept utt2num_frames files to speed up the data loadi…
…ng (#22) * add options to accept utt2num_frames files to speed up the data loading
-
move duplicated network parsers to espresso/speech_tools/utils.py; re…
…move remaining coverage option passed in wsj recipe
-
-
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
-
-
code adaptation/changes according to the commits on Mar 24-Apr 3, 202…
…0; use data.encoders.{bpe,tokenizer} for wordpiece decode -
-
update the qsub script for gpu jobs; code adaptation/changes accordin…
…g to the commits on Apr 16
-
use EncoderOut for SpeechLSTMEncoder's output; code adaptation/change…
…s according to the commits on Apr 21
-
-
-
-
code adaptation/changes according to the commits on Jun 24-25, 2020; fix
validation loss in LSTM models
-
Update Transformer models (#31)
* update transformer * initial recipe * fix transformer * add encoder positional embeddings * add more recipes
-
ignore flake8's FileNotFoundError for soft links to kaldi files; code…
… adaptation/changes according to the commits on Jul 8, 2020
-
fix reorder_encoder_out in SpeechChunkTransformerEncoder; code adapta…
…tion/changes according to the commits on Jul 28, 2020
-
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
-
* 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
-
code adaptation/changes according to the commits on Oct 18-Nov 3, 202…
…0 (lots of changes, mostly for adapting to hydra configs and code formatting)
-
code adaptation/changes according to the commits on Nov 11, 2020; obtain
feat_dim in setup_task() instead
-
code adaptation/changes according to the commits on Nov 16-20, 2020; …
…fix a bug in Multi-level LM when getting cached states
-
fix length tensor device issue in lf_mmi loss; code adaptation/change…
…s according to the commits on Dec 3-12, 2020
