add gh workflow by faroit · Pull Request #70 · sigsep/open-unmix-pytorch (original) (raw)

• Update ptorch 1.5 torchaudio 0.5 (#37)

• update to pytorch 1.5.0 and torchaudio 0.5.0, also remove soundfile loader

• revert conda env naming

• replace soundfile in data test

• move audio loading to data.py

• remove all existing soundfile loading codes

• address comments from @aliutkus

• fix audio loading for soundfile backend

• Add speech enhancement models (#38)

• testing se models on gh

• add README

• fix dataset link, thanks to @TE-StefanUhlich

• End-to-end multichannel wiener filters (#40)

• fix unit tests + regression tests

• update istft with torchaudio.istft

• make inference jitable (check torch.jit.script and torch.jit.trace). Sub-models should be traceble (checked)

• Separator should be scriptable

• test differentiable MWF

• make torchfilters dtype independent

• fix mono support for mwf

• variable renamings

• optimize residual workflow for tensor outputs

• fix silent outputs (happens sometimes when aggregate is active)

• add codecov

• add dep

• add musdb18 tests vectors for spectrogram regression test (#51)

• add musdb18 tests vectors for spectrogram regression test

• save non-centered stft as this is used for training

• update test vectors

• try changing default param

• update env

• update test vectors again – now coming from v1.0.0

• lower tolerance

• lower threshold

• disable codecov for now

• update requirements (#53)

• update requirements

• increase tol

• Revert "increase tol"

This reverts commit 4389f11.

• add bg_iterator

• Add encoder decoder architecture (#52)

• rename targets to target_models

• move model loading to utils

• update batch packing/unpacking from to match torchaudio

• add transform test

• add debug switch for training

• wip

• add first working encoder/decoder

• fix model test

• fix jit test

• fix regression tests

• remove inference test in favor of transforms

• add permute to model to make encoder design cleaner

• fix jit

• pep8

• pep8

• set model specific default targets for separator

• rename functions, create functionality to load separators from torchhub

• update inference docs

• fix evaluation

• rename encoder.json -> separator.json

• Improve deterministic dataset behaviour (#55)

• add deterministic track loading for SourceFolderDataset

• add exceptional treatment for statistics

• glob does not assure deterministic behaviour

• fix train for sourcesfolder

• addressing minor comments from @TE-StefanUhlich

• turn of augmentations by default for more deterministic behavior

• queue iterator prevents deterministic behaviour

• remove random track behaviour

• update sourcefolder dataset to use center segments

• remove comment

• removed unneeded argument

• turn random chunking on by default for sourcefolder dataset

• update training docs

• create dummy dataset and add unit tests

• revert travis activation

• deprecation of sudo=True

• use sh to execute test

• execute tests directly from tracvis

• its a bash script

• speed up test

• Add heterogenous channel datasets (#56)

add stereo_force augmentation and add unit test test for augmentations

• send encoder to cpu for statistics

• update docstrings (#58)

• add mypy annotations

• improve docstrings

• update documentations

• update torch to 1.6.0 (#59)

• update torch to 1.6.0

• add istft functional wrapper for compatibility reasons

• unify backend to support new sox_io backend

• add io unit test to check duration. Also add functionality to discard info loading if already done in preprocessing

• add docstrings

• add test vector

• Add pypi package (#60)

• move into subfolder

• wip

• move hubconf models

• move envs

• change path of env

• contrain requirements

• add envs to scripts

• move train to scripts. make sample rate a float

• rename package to openunmix

• remove temp binary

• finish api

• update dockerfile

• Fix gpu inference (#61)

• fix gpu inference

• also check evaluation and remove numpy support for preprocessing

• re-add shape

• fix training (#62)

• update requirements (#63)

• fix gpu evaluation

• update to museval 0.3.1

• rename functional interface unmix -> separate

• removed predict altogether to add a functional interface in init directly

• remove the reference to predict.inference args

• forgot to remove import of predict

• renamed unmix -> separate

• changed also unmix to separate in cli and setup

• fixed typo

• address some deprecations for torch 1.7.0 (#68)

• address some deprecations

• force 1.7.0 for conda

• replacing torch.atan2 with basic ops to support onnx export

• removing torch.atan2 from wiener function. Magnitude computation to determine cos and sin

• version of wiener using homemade version of atan2

• my_atan2 corrected for case where y=0 and x<0 and for case where x=0 and y=0

• replacing in atan2 numpy.pi with math.pi

• adding rename my_atan2 in atan2, adding docstring, adding torch pi

• disable bg_iterator

• add gh workflow (#70)

• add workflow

• add musdb to tests

• install the right tests

• add ffmpeg

• add museval and musdb

• scipy not needed

• add dummy dataset

• add sox

• add more python version

• add thanks

• add conda test

• typo

• change incubator

• fix python version for conda

• 🐞 maybe its the shell?

• use the incubator workflow

• further simplification

• add libbsndfile

• bye bye travis, I liked you!

• Add stempeg (#71)

• add support for stempeg 0.2.0

• switch to torchaudio 0.7.3

• fix more unit tests

• pass mus object

• drop 1.6.0 support

• update deprecation warning

• Using Asteroid STFTB as an alternative to torch.STFT (#67)

• Using Asteroid STFTB instead of torch.STFT

• bug fix

• adding asteroide to environnements

• correction of merging

• adding Cython in .travis.yml

• using asteroid TorchSTFTFB filterbank wrapped into UMX stft and istft

• Added : lenght as input in asteroid istft and croping the output of stft to fit the original size. Asteroid stft now passes test_regression

• wip : fixing cython build issue

• wip add asteroid-filterbanks in yml file

• handling lenght in istft function

• wip : solving requirement Cython issues

• removing cython dependancy

• wip : minor fixes in yml scripts and travis

• minor argument fixes for asteroid stft

• leaving window in torch type in stft

• pep8

• use asteroid by default. Make ISTFT a moduel

• add asteroid filterbanks

• fix requirements

• add asteroid to training

• shorten tests

• update encoder arch

• transform refactor

• reduce number of tests

• realworld time signature

• prepare filterbank support

• make asteroid_filterbanks optional

• add support for switchable filterbanks

• add asteroid to regression test

• add (failing) onnx test, add tracing test for separator

• make asteroid filterbank optional

Co-authored-by: Fabian-Robert Stöter mail@faroit.com Co-authored-by: Fabian-Robert Stöter faroit@users.noreply.github.com

• update_musdb (#76)

• update musdb and museval requirements

• remove vscode autocomple mess

• add cli test (#74)

• add cli test

• typo

• no token

• add gitpython

• cleanup

• support newest stempeg

• update docker

• update Readme

• fix docstring

• update_docs (#77)

• update readme and training doc

• enhanced docs

• add black/flake8 tests

• add pdoc docs

Co-authored-by: Antoine Liutkus antoine.liutkus@inria.fr Co-authored-by: Baldwin-disso 60886988+Baldwin-disso@users.noreply.github.com