{"id":15066285,"url":"https://github.com/rwth-i6/returnn","last_synced_at":"2025-05-14T13:06:55.838Z","repository":{"id":9192378,"uuid":"61130280","full_name":"rwth-i6/returnn","owner":"rwth-i6","description":"The RWTH extensible training framework for universal recurrent neural networks","archived":false,"fork":false,"pushed_at":"2025-05-08T16:16:48.000Z","size":28898,"stargazers_count":365,"open_issues_count":206,"forks_count":132,"subscribers_count":26,"default_branch":"master","last_synced_at":"2025-05-13T07:57:33.679Z","etag":null,"topics":["deep-learning","gpu","recurrent-neural-networks","tensorflow","theano"],"latest_commit_sha":null,"homepage":"http://returnn.readthedocs.io/","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"other","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/rwth-i6.png","metadata":{"files":{"readme":"README.rst","changelog":"CHANGELOG.md","contributing":"CONTRIBUTING.md","funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":"CODEOWNERS","security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null}},"created_at":"2016-06-14T14:37:14.000Z","updated_at":"2025-05-11T01:24:46.000Z","dependencies_parsed_at":"2023-09-23T07:17:41.599Z","dependency_job_id":"94e4a71a-59d1-4d14-a00f-223b804d14a6","html_url":"https://github.com/rwth-i6/returnn","commit_stats":{"total_commits":11350,"total_committers":91,"mean_commits":"124.72527472527473","dds":"0.27004405286343613","last_synced_commit":"ed5db537bc10999c4b1cb2b84c3c73a28db4ab85"},"previous_names":[],"tags_count":3,"template":false,"template_full_name":null,"repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rwth-i6%2Freturnn","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rwth-i6%2Freturnn/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rwth-i6%2Freturnn/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rwth-i6%2Freturnn/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/rwth-i6","download_url":"https://codeload.github.com/rwth-i6/returnn/tar.gz/refs/heads/master","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":254149955,"owners_count":22022851,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["deep-learning","gpu","recurrent-neural-networks","tensorflow","theano"],"created_at":"2024-09-25T01:05:04.640Z","updated_at":"2025-05-14T13:06:50.817Z","avatar_url":"https://github.com/rwth-i6.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"==================\nWelcome to RETURNN\n==================\n\n`GitHub repository \u003chttps://github.com/rwth-i6/returnn\u003e`__.\n`RETURNN paper 2016 \u003chttps://arxiv.org/abs/1608.00895\u003e`_,\n`RETURNN paper 2018 \u003chttps://arxiv.org/abs/1805.05225\u003e`_.\n\nRETURNN - RWTH extensible training framework for universal recurrent neural networks,\nis a Theano/TensorFlow-based implementation of modern recurrent neural network architectures.\nIt is optimized for fast and reliable training of recurrent neural networks in a multi-GPU environment.\n\nThe high-level features and goals of RETURNN are:\n\n* **Simplicity**\n\n  * Writing config / code is simple \u0026 straight-forward (setting up experiment, defining model)\n  * Debugging in case of problems is simple\n  * Reading config / code is simple (defined model, training, decoding all becomes clear)\n\n* **Flexibility**\n\n  * Allow for many different kinds of experiments / models\n\n* **Efficiency**\n\n  * Training speed\n  * Decoding speed\n\nAll items are important for research, decoding speed is esp. important for production.\n\nSee our `Interspeech 2020 tutorial \"Efficient and Flexible Implementation of Machine Learning for ASR and MT\" video \u003chttps://www.youtube.com/watch?v=wPKdYqSOlAY\u003e`__\n(`slides \u003chttps://www-i6.informatik.rwth-aachen.de/publications/download/1154/Zeyer--2020.pdf\u003e`__)\nwith an introduction of the core concepts.\n\nMore specific features include:\n\n- Mini-batch training of feed-forward neural networks\n- Sequence-chunking based batch training for recurrent neural networks\n- Long short-term memory recurrent neural networks\n  including our own fast CUDA kernel\n- Multidimensional LSTM (GPU only, there is no CPU version)\n- Memory management for large data sets\n- Work distribution across multiple devices\n- Flexible and fast architecture which allows all kinds of encoder-attention-decoder models\n\nSee `documentation \u003chttps://returnn.readthedocs.io/\u003e`__.\nSee `basic usage \u003chttps://returnn.readthedocs.io/en/latest/basic_usage.html\u003e`__\nand `technological overview \u003chttps://returnn.readthedocs.io/en/latest/tech_overview.html\u003e`__.\n\n`Here is the video recording of a RETURNN overview talk \u003chttps://www-i6.informatik.rwth-aachen.de/web/Software/returnn/downloads/workshop-2019-01-29/01.recording.cut.mp4\u003e`_\n(`slides \u003chttps://www-i6.informatik.rwth-aachen.de/web/Software/returnn/downloads/workshop-2019-01-29/01.returnn-overview.session1.handout.v1.pdf\u003e`__,\n`exercise sheet \u003chttps://www-i6.informatik.rwth-aachen.de/web/Software/returnn/downloads/workshop-2019-01-29/01.exercise_sheet.pdf\u003e`__;\nhosted by eBay).\n\nThere are `many example demos \u003chttps://github.com/rwth-i6/returnn/blob/master/demos/\u003e`_\nwhich work on artificially generated data,\ni.e. they should work as-is.\n\nThere are `some real-world examples \u003chttps://github.com/rwth-i6/returnn-experiments\u003e`_\nsuch as setups for speech recognition on the Switchboard or LibriSpeech corpus.\n\nSome benchmark setups against other frameworks\ncan be found `here \u003chttps://github.com/rwth-i6/returnn-benchmarks\u003e`_.\nThe results are in the `RETURNN paper 2016 \u003chttps://arxiv.org/abs/1608.00895\u003e`_.\nPerformance benchmarks of our LSTM kernel vs CuDNN and other TensorFlow kernels\nare in `TensorFlow LSTM benchmark \u003chttps://returnn.readthedocs.io/en/latest/tf_lstm_benchmark.html\u003e`__.\n\nThere is also `a wiki \u003chttps://github.com/rwth-i6/returnn/wiki\u003e`_.\nQuestions can also be asked on\n`StackOverflow using the RETURNN tag \u003chttps://stackoverflow.com/questions/tagged/returnn\u003e`_.\n\n.. image:: https://github.com/rwth-i6/returnn/workflows/CI/badge.svg\n    :target: https://github.com/rwth-i6/returnn/actions\n\nDependencies\n============\n\npip dependencies are listed in ``requirements.txt`` and ``requirements-dev``,\nalthough some parts of the code may require additional dependencies (e.g. ``librosa``, ``resampy``) on-demand.\n\nRETURNN supports Python \u003e= 3.8. Bumps to the minimum Python version are listed in `CHANGELOG.md \u003chttps://github.com/rwth-i6/returnn/blob/master/CHANGELOG.md\u003e`__.\n\nTensorFlow-based setups require TensorFlow \u003e= 2.2.\n\nPyTorch-based setups require Torch \u003e= 1.0.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Frwth-i6%2Freturnn","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Frwth-i6%2Freturnn","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Frwth-i6%2Freturnn/lists"}