Dummy 2 (#227)

* Add swb PyTorch ctc setup (#219)

* add initial setup

* rm binary files

* rm binary files

---------

Co-authored-by: Jingjing Xu <[email protected]>

* add conformer enc with more weight dropout

* fix

* add more weight noise opts

* black formatting

* Jing tedlium independent softmax (#220)

* tedlium ctc pytorch

* rm empty files

---------

Co-authored-by: Jingjing Xu <[email protected]>

* use ff regs for mhsa out

* add more regularized trafo dec

* update

* add regs to rnn decoder

* more

* add more regs to rnn dec

* black formatting

* Update users/berger

* update

* add readme for RF

* more

* cleanup, generalize, different spm vocab sizes

* more

* more

* small fix

* update ls att+ctc+lm

* add args

* fix pretraining

* Glow-TTS-ASR: Update with fixed invertibility tests

* Glow-TTS-ASR update

* fix

* Glow-TTS-ASR: Cleanup and comments/documentation

* small fixes

* spm20k

* more ctc and rnn-t librispeech experiments

* add greedy decoder

* black

* add ebranchformer

* more

* more

* better layer names for ebranchformer

* better

* better

* decouple mhsa residual

* cleanup

* refactor args. add ebranch config

* more

* Update users/berger

* Update users/berger

* better

* config enable write cache manager

* standalone 2024 setup add LSTM lm pipeline

* update

* add horovod to libri pipeline

* update configs

* fix

* more

* update

* update zoneout fix ted2

* Update users/berger

* Update users/berger

* more

* update

* update

* Update users/berger

* more

* more

* update

* update

* ConformerV2 setup

* updates

* cleanup

* updates and fix mel norm + zoneout

* update conf v2

* more

* update

* Update users/berger

* fixes and update

* add CTC gauss weights

* convert ls960 LSTM LM to rf

* fix

* fix

* fix

* more

* more

* use_eos_postfix

* fix CTC with EOS recog scoring

* ctc eos fix more

* update

* Update Glow-TTS-ASR

* updates quant

* added factored bw

* deleted wrong stashed

* update trainings and initial rnnt decoder rf

* feature batch norm

* feature normalization

* recog fix API doc

* collect stats, initial code

* librispeech feature stats

* feature global norm

* small fixes

* small fix

* small fix

* small fix

* fix feat norm search

* update

* update

* update

* add more weight drop to rnn decoder

* add chunked rnn decoder

* update

* fix

* fix

* more

* more

* fix name

* more

* more

* small fix

* more

* more

* cleanup

* fix

* comment

* more

* cleanup

* more

* add canary 1b recog sis prepare config

* add config (#223)

Co-authored-by: Jingjing Xu <[email protected]>

* more

* add nemo model download job

* add nemo search job

* add custom hash

* fix

* add nemo search

* first version of nemo search

* better

* fix bug

* better

* add missing search output path

* add compute_wer func

* add wer as output var

* run search for all test sets with canary 1b model

* add configs (#224)

Co-authored-by: Jingjing Xu <[email protected]>

* update

* update

* register wer as out

* update

* add libri test other test set

* fix args

* fix args

* update

* add modified normalized

* Create README.md

* Update README.md

* Update users/berger

* Update users/berger

* more

* more

* more

* more

* more

* more

* prepare for some more modeling code

* move SequentialLayerDrop

* better

* move mixup

* rnnt dec rf WIP

* Update users/berger

* update users/raissi monofactored

* update

* update

* ls960 pretrain: use phoneme info for mask boundaries

* BatchRenorm initial implementation (untested)

* test_piecewise_linear

* test_piecewise_linear use dyn_lr_piecewise_linear

* dyn_lr_piecewise_linear use RETURNN PiecewiseLinear

* DeleteLemmataFromLexiconJob (#225)

* ls960 pretrain: phoneme mask and other targets

* ls960 pretrain: update num epochs

* better

* first version of beam search

* fix

* fix enc shape

* use expand instead of repeat for efficiency

* better

* add hyp postprocessing

* better

* add beam search

* remove print

* more

* BatchRenorm with build_from_dict

* more

* small fix

* more

* small fix

* reorder code

* comment

* prior

* cleanup

* cache enc beam expansion

* fix bug

* update

* more

* more

* more

* more

* LS spm vocab alias

* make private

* move

* lazy, aliases

* update and test rf vs torch mhsa

* fix warning

* fix bug

* vocab outputs

* more

* more, AED featBN, sampling

* extract SPM vocab

* add rtfs

* add cache suffix

* update

* fix

* add debug out

* add batch size logging

* import i6_models conformer in rf, batch 1

* SamplingBytePairEncoding for SentencePiece

* add gradient clipping to example baseline

* 2-precision WER and quantization helper

* HDF alignment labels example data pipeline

* ls960 pretrain: fix python launcher for itc/i6

* latest users/raissi

---------

Co-authored-by: Judyxujj <[email protected]>
Co-authored-by: Jingjing Xu <[email protected]>
Co-authored-by: Mohammad Zeineldeen <[email protected]>
Co-authored-by: Simon Berger <[email protected]>
Co-authored-by: schmitt <[email protected]>
Co-authored-by: Albert Zeyer <[email protected]>
Co-authored-by: luca.gaudino <[email protected]>
Co-authored-by: Lukas Rilling <[email protected]>
Co-authored-by: Nick Rossenbach <[email protected]>
Co-authored-by: Benedikt Hilmes <[email protected]>
Co-authored-by: Mohammad Zeineldeen <[email protected]>
Co-authored-by: Peter Vieting <[email protected]>
Co-authored-by: vieting <[email protected]>

Author: 13 people