Upgrade to Torch 2.x and Lightning 2.x

⬆️ Upgrade ⬆️

Changes documented in #682 :

• Remove the optimizer_idx in lr_scheduler_step in System.
• Replace torch.symeigh by torch.linalg.eigh in beamforming.py
• Disable X-UMX as it uses torch1.x STFT : @r-sawata do you want to fix it ?
• Replace on_*_end with on_training_*_end
• Replace torch.testing.assert_allclose with torch.testing.assert_close.

Happy coding 🙃

[Summer beauty] Prepare upgrade to Torch 2.x and Lightning 2.x

Release before moving to asteroid 0.7.x with torch and lightning upgrades.

Breaking

• [install] Limit torch and pytorch-lightning to 1.x (#671)

Added

• [egs] Raise error if storage_dir not specified in LibriMix recipes (#626)
• [egs] Create Multi-Decoder DPRNN README.md (#632)
• [egs] Update Multi-Decoder DPRNN with inference function on a single file out-of-the-box (#653)
• [egs] Add a new pre-trained model (X-UMXL) (#665)
• [egs] Add a notebook demo, separate.py script for MD-DPRNN. Easier use from_pretrained (#668)

Changed

• [src] Use torch.linalg for solve and cholesky (#623)
• [install] Remove torchmetrics version limit
• [CI] Lint with black 22.3.0 and fix linting CI python3.*8 (#624)

Fixed

• [hub] Upgrade deprecated huggingface_hub.cached_download (#645)
• [src] Fix cumulative normalization (#649)
• [src] Fix argparse key based on Python version (#628) (#657)
• [ci] Fix tests in continuous integration (#672)

Many thanks to all the contributors @actuallyaswin, @mystlee, @JunzheJosephZhu, @jbartolewska, @LeonieBorne, @mattiadg, @r-sawata and @zmolikova ! 💪 🤩 🙏

Upgrade to newer PyTorch-Lightning versions (1.5+)

Yes. That's it. Only Pytorch-Lightning support.

From 1.5.0 to the latest versions.

So changing to the new Trainer API with the new accelerator, strategy and devices arguments.

Thanks to @paulfd to push me to do it 💪

Small patch release

Minor patch release before 0.6.0 that will upgrade pytorch-lightning version.

Added

• [egs&tests] MixIT loss function (#595)

Changed

• [docs] Update RTD version to 0.5.1

Fixed

• [src] Fix FasNetTAC loading (#586)
• [src] Fix device in padding for sudormrf. Fix #598 (#603)
• [docs] Fix deep_clustering_loss docs example (#607)
• [src] Remove torch.complex32 usage for torch 1.11.0 (#609)
• [egs] Fix package install script for WHAMR! (#613)
• [src] Fix shape mismatching in SuDORMRF's masknn (#618)
• [CI] Fix CI Restrict torchmetrics version to under 0.8.0 (#619)
• [docs] Fix docs Restrict jinja2>=3.0.0,<3.1.0 (#620)

Thanks to @jc5201, @z-wony, @JorisCos, @ben-freist, @nicocasaisd and @zmolikova for their awesome contributions 🔥 💪 🙏

First Asteroid-based VAD

Highlights

• New ConvTasNet-based VAD on the Libri_VAD dataset 🚀
• Fixed and improved GEVD beamformer
• Multi-Decoder DPRNN recipe

Changelog

Added

• [src&egs] Add VAD in Asteroid 🎉 (#558)
• [src&tests] Add GEVD beamformer (#520)
• [hub] Add library version and name for Hugging Face download stats (#524)
• [egs] Add recipe for Multi-Decoder DPRNN (#463)
• [egs] Enable WER eval with GPU (#541)
• [egs] Add readme and pretrained model to FaSNet recipe (#561)

Changed

• [src] Make cLN default in causal ConvTasNet (#511)
• [install] Pin pytorch-optimizer version to support RAdam (#568)

Fixed

• [src] Prevent complex eigen values in gev (#519)
• [src] Fix default norm_type for causal ConvTasNet (#503)
• [egs] Fixed bug in X-UMX (#521)
• [nbs] asteroid.filterbanks to asteroid_filterbanks (#526)
• [nbs] Fix typos in notebooks/02_Filterbank.ipynb (#527)
• [install] Fix PL version <1.5.0 (#576)

Thanks to @ldelebec @hihunjin @nobel861017 @ben-freist @r-sawata @osanseviero @JunzheJosephZhu and @JorisCos for their awesome contributions 🔥 💪 🙏

PyTorchLightning1.3.0 support

Nothing more, nothing less 🙃
Check the release below for the most recent release notes.

Torch>=1.8 only, better beamforming and MDX challenge baseline

⚠️ This release drops support for PyTorch under 1.8 and restricts PyTorch-Lightning under 1.3.
The next release (0.5.1) will add support to PyTorch-Lightning 1.3.0 which broke our CI. ⚠️

Highlights

• Beamforming modules are now more stable (computed in double by default), have been renamed and extended.
• The baseline for the Music Demixing Challenge has been released 🎉

Changelog

Breaking

• [src&tests] Remove all deprecated code (#474)
• [all] Drop support for torch<1.8.0 (#476)

Added

• [src] Beamforming: Souden MVDR and optimal channel selection (#484)
• [src&egs] X-UMX Official baseline for Music Demixing Challenge (#490)

Changed

• [src] Compute linalg beamforming ops in double (#482)
• [src] Improve Beamforming naming and add TODOs (#483)
• [src] Beamforming: enable forcing float linalg ops (#485)
• [docs] Update instructions for pretrained model sharing (#489)
• [install] Bump lightning version under 1.3.0 (#493)

Fixed

• [nb] Fix shape in 00_GettingStarted.ipynb (#478)
• [src] Stabilize GEV beamformer (#479)
• [src] Beamforming: fix doc references

Thanks to @quancs @r-sawata @popcornell and @faroit for their contributions

Beamforming modules and last torch<1.8 release

⚠️ Warning ⚠️ this is the last release that supports torch<1.8.

From asteroid 0.5.0 onwards, only torch>=1.8.0 will be supported. The main reason being the complex support, the fft and linalg packages.

Highligths

• We now have end to end beamforming modules! 🎉 Multichannel recipes will follow.

Changelog

• [src] Fix padding in DCUNet ConvTranspose2d (#466)
• [src&tests] Causal TDConvNet and ConvTasNet (#465)
• [tests] Fix test from #465 (#467)
• [docs] Fix wrong doclinks (#469)
• [docs] Fix docs (links and errors) (#470)
• [src&tests] Add beamforming modules 🎉 (#468)

Large thanks to the contributors ! 🙃

Builtin multichannel model & LibriMix recipes

Highlight

• Deprecated: Not feeding sample_rate to BaseModel is deprecated, and it will raise an error in the future release.
• BaseModel now takes a in_channels argument which will be used in separate and the asteroid-infer CLI.
  This makes it now possible to share multichannel models on the hub (for example)
• The first fully supported multichannel model is FasNetTAC thanks to @popcornell ! 🎉
• Use huggingface_hub instead of "our own" code for interfacing with the Hub.

Changelog

Breaking

• [src] Make sample_rate positional in BaseModel (#431)

Added

• [src&egs] ESPNet in LibriMix recipe 🎉 (#329)
• [cli] Add --device to asteroid-infer (#375)
• [src] Add stft_n_filters to BaseDCUNet (#406)
• [src&tests] Add MetricTracker (#394)
• [egs] Librimix recipes for all models (#418)
• [src] Track transcriptions in WerTracker (#414)
• [docs] Add note about System Lightning hooks (#428)
• [src] Support Multichannel models (#427)
• [src&egs] Add FasNetTAC model, dataset and recipe (#306)
• [src] Add mulcat option to DPRNN (#416)

Changed

• [src&install] Drop librosa and refactor requirements files (#386)
• [src] Remove unused hooks (#424)
• [hub] Use huggingface_hub deps, remove inlined HF code (#409)

Fixed

• [egs] Pass samplerate to model (#407)
• [src] Fix Large-DCUNet-20 arch (#405)
• [src] Fix inconsistent devices for tensors (#417)
• [egs] Fix model saving path in DeepClustering recipes(#398)
• [src] Fix unpassed SR in TasNet and SudoRMRF (#433)
• [egs] Fix reshaping issue in AVSpeech (#441)

Lightning1.x upgrade and HuggingFace's model hub

Highlights

• Finally upgrade to Lightning 1.x. Have a look at their docs to see what's new! In particular, callbacks and logging have changed a bit, for the better 😉
• Migration from the Zenodo hub to the HuggingFace's model hub, we even have a tag and a Use in Asteroid button, this is great! 🤩 Huge thanks to the HuggingFace team and @julien-c in particular. 🙏

Changelog

Added

• [hub] Support for huggingface model hub 🎉 (#377)
• [hub] List asteroid models HF's hub (#382)
• [install] Deduplicate version number (#388)

Changed

• [docs] Migrate to asteroid-team (#369)
• [src] Import BaseModel from .models
• [src] Upgrade Lightning to 1.x (#371)

Fixed

• [install] Fix old STFT model loading (thanks @popcornell)
• [hub] Fix torch.hub tests Add new dependencies from #377
• [src] bug in attention.py due to mha input shape (#381)
• [src] Hopefully fixes DPTNet this time (#383)
• [src] Remove call to super().training_step() (#395)

Thanks to all the contributors: @jonashaag, @popcornell, @julien-c, @iver56, @lubacien, @cliffzhao and the issue creators and bug-reporters 🙏