Datasets:

---

renwei2024

/

nemo-github-issues

like 0

[]

[]

[ "Could you please sign your commits. See https://github.com/NVIDIA/NeMo/blob/main/CONTRIBUTING.md for general guidelines.\r\nTo sign already committed commits, checkout DCO's bot output: https://github.com/NVIDIA/NeMo/pull/1678/checks?check_run_id=1789269686", "CI failed on style check. \r\nTo check for style do:\r\n``python setup.py style``\r\nTo fix style:\r\n``python setup.py style --fix``", "This pull request **introduces 4 alerts** when merging 057c428dde16e9bb62a5dafaf1565378047cb733 into 8a26496f415386f50adcead21ebd1e4a21a7ff42 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-876474ea51b450756c5d62c7fafd1b354a66b534)\n\n**new alerts:**\n\n* 4 for Unused import", "This PR requires a significant refactor - and does not need to add explicit global support for quantization in the constructor. If you would agree to it, could I push some changes to this branch directly ?", "> This PR requires a significant refactor - and does not need to add explicit global support for quantization in the constructor. If you would agree to it, could I push some changes to this branch directly ?\r\n\r\nThanks @titu1994 , I do not know other ways to pass the quantization flag from the upper script to model, please help me to refactor the code, thanks!", "Sounds good, thanks for the go ahead.", "@okuchaiev for review regarding core ModelPT changes to support 1) returning of config instead of model 2) overriding config with either str path or with OmegaConf object", "This pull request **introduces 4 alerts** when merging de1f57db1a375b8a6972b07d5c870c7ec6467fc0 into 8a26496f415386f50adcead21ebd1e4a21a7ff42 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-7df2249de95d719f81c51b4c2cbecf9f9508e2bf)\n\n**new alerts:**\n\n* 4 for Unused import", "This pull request **introduces 4 alerts** when merging 033b96296c5b72f82491fda49feb0cd03364c27f into 8a26496f415386f50adcead21ebd1e4a21a7ff42 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-ef25758a3207343c34dd3975db75604fa7f111b5)\n\n**new alerts:**\n\n* 4 for Unused import", "This pull request **introduces 4 alerts** when merging 33d0c11416a0b2b02b8406a02bddb9b1dfd5490e into 8a26496f415386f50adcead21ebd1e4a21a7ff42 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-dab7b7231360f2ca3eacc5b1952a1d6341c63373)\n\n**new alerts:**\n\n* 4 for Unused import", "This pull request **introduces 4 alerts** when merging e542e45a3c616fa1b5b03ba30d1768df148bd9a5 into 8a26496f415386f50adcead21ebd1e4a21a7ff42 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-dd5e821caa2ca951023a3d81be5b4b2f4dffaa53)\n\n**new alerts:**\n\n* 4 for Unused import" ]

[]

[]

[ "Every NeMo model inherits from ``torch.nn.Module`` at the top of it's inheritance chain. So everything applicable to torch.nn.Module objects, including serving them via Triton Server should apply here as well.\r\n\r\n", "Great, thanks for the response. " ]

[]

[]

[ "Hi, is there any update on this?\r\n\r\nI would also like to ask if it's possible to add a language model in this notebook :\r\nhttps://colab.research.google.com/github/NVIDIA/NeMo/blob/main/tutorials/asr/01_ASR_with_NeMo.ipynb\r\n\r\nI would like to add my language model to train my data in order to achieve better WER, so far I could not find any tutorial on this.\r\n\r\nLooking forward for your replies,\r\n\r\nThanks.\r\n\r\n\r\n\r\n\r\n\r\n\r\n", "Documentation for language modelling with ASR models is available here - https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html" ]

[]

[]

[]

[]

[]

[ "Thanks @yzhang123 !" ]

[]

[]

[ "This pull request **introduces 4 alerts** when merging 8804196b70e585ebb8bd2847f4a503ba19358ac2 into cd8f1c5a74e0d0f2c5bf5847951e30d1170909c3 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-33c1b211c8938e3824030f9c8b2b1d9708c3bfdb)\n\n**new alerts:**\n\n* 3 for Unused import\n* 1 for Unused local variable", "This pull request **introduces 4 alerts** when merging e6d59380c7f599616d043750c4d58eefd001a7f2 into cd8f1c5a74e0d0f2c5bf5847951e30d1170909c3 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-b707f2a3fe60823dab01087fadc6822aac488bf3)\n\n**new alerts:**\n\n* 4 for Unused import", "This pull request **introduces 1 alert** when merging 8d0d83dd2b36e1d250f2aacbb63471b7dbc00127 into cd8f1c5a74e0d0f2c5bf5847951e30d1170909c3 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-eb9c72b8bb41b91e29b388e0b4010687909a0fb5)\n\n**new alerts:**\n\n* 1 for Unused import", "@MaximumEntropy Try the final dataloader on a multigpu setup to full convergence before merge. Otherwise, it looks good. ", "this should go to \"main\" (I should probably delete nmtmodel branch). And now there are a bunch of merge conflicts. Any thoughts why?", "@okuchaiev I guess its because this branch was created from nmtmodel rather than main and there have been some commits to main after the nmtmodel merge (like typing). I can either resolve these conflicts or just submit another PR starting from main.", "Closing this PR since it started from the wrong branch. New one at #1683 " ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[]

[ "can you please send a PR?", "fixed" ]

[]

[]

[]

[]

[]

[ "@okuchaiev Could you review the move of citrinet configs from experimental to conf/citrinet/ and update to Jenkinsfile? " ]

[]

[]

[ "This pull request **introduces 1 alert** when merging 03dc67e4e06a7fd939ec32ae6a3ae2ed07371759 into ad0edf2e6910b7b6de3ae93fba9b728f6479c50e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-c00b16affad2c7a4c82e0998d3d0995a058cd2fa)\n\n**new alerts:**\n\n* 1 for Unused import", "This pull request **introduces 1 alert** when merging aef07cbe74e840895837e762d43cec01edf20520 into ad0edf2e6910b7b6de3ae93fba9b728f6479c50e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-9cf8fbb6a2d4f0b68c50296c181384e3550def58)\n\n**new alerts:**\n\n* 1 for Unused import", "This pull request **introduces 4 alerts** when merging 8c31f4a4477ff00015b8ae2090088037efdc4117 into ad0edf2e6910b7b6de3ae93fba9b728f6479c50e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-3cd0f71bd1172d2ed2de884efa167a1af42e6667)\n\n**new alerts:**\n\n* 4 for Unused import", "This pull request **introduces 6 alerts** when merging 02207f86415e790202848f2c36cec7b9c4d434d5 into ad0edf2e6910b7b6de3ae93fba9b728f6479c50e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-c9c4ab115bd8a7f9e0c5fcea57199420d23b2096)\n\n**new alerts:**\n\n* 6 for Unused import", "This pull request **introduces 3 alerts** when merging d5d4b232bad996a0d580cf60861da375c6e67d94 into ad0edf2e6910b7b6de3ae93fba9b728f6479c50e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-8034262612377c60fb47429f0eaed321a4ed596c)\n\n**new alerts:**\n\n* 3 for Unused import", "This pull request **introduces 6 alerts** when merging 8035fb988a2ea99060906d61e556c4c567f0783e into ad0edf2e6910b7b6de3ae93fba9b728f6479c50e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-2a53970fd049a53b63fe0aecc48e2ebba93538d3)\n\n**new alerts:**\n\n* 6 for Unused import", "This pull request **introduces 6 alerts** when merging 13e9523583dee6c3162f788f0ded35d40e8be195 into ad0edf2e6910b7b6de3ae93fba9b728f6479c50e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-8a488418abeec43276230b5d76f0919febc82bbb)\n\n**new alerts:**\n\n* 6 for Unused import", "This pull request **introduces 7 alerts** when merging c8cab3bb299ec160506d29049d535b204461ce7a into 838892b1d38650bdf1bd829437f5722589f3ae42 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-0e82ea57b3b8d1489dc063ca5698a74d96827759)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **introduces 7 alerts** when merging d140ecb485c0e6ee39187dbde63c78c7c6b578ff into c4379da172c1ac866fea0cd91ac6433dcd119651 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-cd9d7dcd746bf7bc876c7b0f2d4fdd66b0034e4d)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **introduces 7 alerts** when merging 4f612b8e057aad68b38eb0640de1fab9ee892412 into 4fe96b152a74ba2dd3e044cd76fe91dfda566c08 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-be0ac3f1ed9da68c4be76cfc657b8496201aaa83)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **introduces 7 alerts** when merging 26e7446e098ed675511db01582d6d0aabd0bedc0 into 3a67cd013c451f65eba75a0c54dcdcad89c9e95e - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-ddc1888fd8d6a12ea682855215b9eb37b49fe0ce)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **introduces 7 alerts** when merging 1e1e08a9c0e991ccf20bce8e71d9b394065aa068 into 4fc670ad0c886be2623247921d4311ba30f486f8 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-6201fb13723cee2bb0b099feab30fc972ec9dc45)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **introduces 7 alerts** when merging 1af04e4192fbf8b240cbc8a7e6924619b2479397 into 4fc670ad0c886be2623247921d4311ba30f486f8 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-ed7ec377da0e2e12033f12e9d30436abf85bb56b)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **introduces 7 alerts** when merging 31b2529424ef89aa61cd21bd24420539ce5bac84 into 4fc670ad0c886be2623247921d4311ba30f486f8 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-65d9475e9b65afc5f4b6358083158f4bc2d5bc0d)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **introduces 7 alerts** when merging 20a37cec79729323ba976a0905e37daa019ea1f9 into 4fc670ad0c886be2623247921d4311ba30f486f8 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-3b89e3a9955ae19480c4127dd192c30b4c4d4048)\n\n**new alerts:**\n\n* 7 for Unused import", "This pull request **fixes 2 alerts** when merging 0552327ff2bb3f84eec23a9643a50c8c9b3525f8 into 4fc670ad0c886be2623247921d4311ba30f486f8 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-c5bcb22771d638b0df7c818b0e55893454023d6b)\n\n**fixed alerts:**\n\n* 2 for Unused import", "This pull request **fixes 2 alerts** when merging 6da45cfd3132804ede19bebde0e22479b3640619 into 6c34fe9d44d0c398dd9a87404f9fa45616a13647 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-50cdc7a69a1c92dbc3d8880909231c3d36f5252d)\n\n**fixed alerts:**\n\n* 2 for Unused import", "This pull request **fixes 2 alerts** when merging 4e3fdaac7c945785149d74d74b0f7a41cdc03451 into 3c9e60ab269e858cc73e780e5f85f749a8b8dcff - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-03fbce0110733279aa8078fea9fdb5858ebd1f55)\n\n**fixed alerts:**\n\n* 2 for Unused import", "This pull request **fixes 2 alerts** when merging a6f200f29fe688b91db0cf82565542d1fcd5cb50 into 090cfae5db9b6f9eb62b3ede37ed438875a150df - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-c28ba7668fcf2c5ae9052379e659cd92127064e0)\n\n**fixed alerts:**\n\n* 2 for Unused import", "This pull request **fixes 2 alerts** when merging 65e1b3f20fa063206b8cdd4ff043572a19f55fa3 into 090cfae5db9b6f9eb62b3ede37ed438875a150df - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-bc0906d35565b3d2304aa6c1589223d1e1e5b2c0)\n\n**fixed alerts:**\n\n* 2 for Unused import", "This pull request **fixes 2 alerts** when merging 368d32bfbbe945426ce30025692506a80061e3ca into 4f87abb242fee7f0131fd27a266883c8740d8fac - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-d83fa003be981db281e39d1f6a282e677680386a)\n\n**fixed alerts:**\n\n* 2 for Unused import" ]

[]

[]

[ "Currently we do not have wav2letter implementation. Checkout Jasper and QuartzNet networks which were inspired by wav2letter", "Ok, thank you." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "This pull request **introduces 1 alert** when merging cdb20c1742dca80fd3519991029d356e95573f98 into 9ea071605034f9f56db9b677eb1fe54fdee2b820 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-cc6090175340df7c3bed280609bfe623e3253748)\n\n**new alerts:**\n\n* 1 for Unused import", "This pull request **introduces 1 alert** when merging 6ad69e3b240fe8ffedcaae7d4260dcbd05828c1b into 9ea071605034f9f56db9b677eb1fe54fdee2b820 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-2bbd5a3297d9a000e28efc1a91aba9366c42046e)\n\n**new alerts:**\n\n* 1 for Unused import" ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[ { "login": "ekmb", "id": 10428420, "node_id": "MDQ6VXNlcjEwNDI4NDIw", "avatar_url": "https://avatars.githubusercontent.com/u/10428420?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ekmb", "html_url": "https://github.com/ekmb", "followers_url": "https://api.github.com/users/ekmb/followers", "following_url": "https://api.github.com/users/ekmb/following{/other_user}", "gists_url": "https://api.github.com/users/ekmb/gists{/gist_id}", "starred_url": "https://api.github.com/users/ekmb/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ekmb/subscriptions", "organizations_url": "https://api.github.com/users/ekmb/orgs", "repos_url": "https://api.github.com/users/ekmb/repos", "events_url": "https://api.github.com/users/ekmb/events{/privacy}", "received_events_url": "https://api.github.com/users/ekmb/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[]

[]

[]

[ "This pull request **introduces 6 alerts** and **fixes 1** when merging af4278cd307544343b466d80f681f939f41d7a3e into df360cd3c856c645d23505305ee8b689636b2dd1 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-976deade5b08d9a975c92bc6433f92edeaf9f4ab)\n\n**new alerts:**\n\n* 3 for Except block handles 'BaseException'\n* 3 for Unused import\n\n**fixed alerts:**\n\n* 1 for Except block handles 'BaseException'", "This pull request **introduces 1 alert** when merging 9ed7a43b4919facffeee454d1fe3ff26bc182d17 into 9ea071605034f9f56db9b677eb1fe54fdee2b820 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-798486a6838f30a2b5a1dd7ef2a11ee73916b0c5)\n\n**new alerts:**\n\n* 1 for Unused import" ]

[ { "id": 1719393562, "node_id": "MDU6TGFiZWwxNzE5MzkzNTYy", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/feature", "name": "feature", "color": "68F3F0", "default": false, "description": "request/PR for a new feature" } ]

[ { "login": "okuchaiev", "id": 22035961, "node_id": "MDQ6VXNlcjIyMDM1OTYx", "avatar_url": "https://avatars.githubusercontent.com/u/22035961?v=4", "gravatar_id": "", "url": "https://api.github.com/users/okuchaiev", "html_url": "https://github.com/okuchaiev", "followers_url": "https://api.github.com/users/okuchaiev/followers", "following_url": "https://api.github.com/users/okuchaiev/following{/other_user}", "gists_url": "https://api.github.com/users/okuchaiev/gists{/gist_id}", "starred_url": "https://api.github.com/users/okuchaiev/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/okuchaiev/subscriptions", "organizations_url": "https://api.github.com/users/okuchaiev/orgs", "repos_url": "https://api.github.com/users/okuchaiev/repos", "events_url": "https://api.github.com/users/okuchaiev/events{/privacy}", "received_events_url": "https://api.github.com/users/okuchaiev/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "This is not a good approach imo, because it allows the chance that model weights are skipped with a warning only and training progresses as is. When developing on scheduled dispatchers like Slurm, we rarely every take constant look at slurm logs so this can silently introduce bugs.\r\n\r\nThe official way to change the decoder is using the provided method - it contructs a new randomly initialized decoder for you. If you only wish to load encoder weights, that is possibly as below\r\n\r\n```python\r\nold_ckpt = torch.load(...)\r\nnew_model = EncDecCTCModel(...)\r\n\r\n# change decoder vocabular\r\nnew_model.change_vocabulary(...)\r\n\r\n# do things manually (NOT ADVISED)\r\nnew_model.decoder = torch.nn.Sequential(...)\r\nnew_model.encoder.load_state_dict(old_ckpt['state_dict']['encoder'])\r\n...\r\n```", "Thank you for the response! Yeah, that's a fair point. I'll just stick with the provided method then." ]

[]

[]

[ "1. NeMo software is released under [Apache 2.0 license](https://github.com/NVIDIA/NeMo/blob/main/LICENSE)\r\n2. Pre-trained checkpoints which you can find on NVIDIA NGC are under [NGC TERMS OF USE](https://ngc.nvidia.com/legal/terms) unless another License/Terms Of Use/EULA is clearly specified.\r\n\r\nSpecifically, all models from:\r\n\r\n- https://ngc.nvidia.com/catalog/models/nvidia:nemospeechmodels/files?version=1.0.0a5\r\n- https://ngc.nvidia.com/catalog/models/nvidia:nemonlpmodels/files?version=1.0.0a5\r\n- https://ngc.nvidia.com/catalog/models/nvidia:nemottsmodels/files?version=1.0.0a5\r\n\r\n(note ``?version=1.0.0a5`` part) are under [NGC TERMS OF USE](https://ngc.nvidia.com/legal/terms).\r\n\r\nI think going forward we should explicitly specify License/EULA for pre-trained models' weights. Will try doing that in the next releases.\r\n\r\nWe are also open to community's feedback on this question." ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[ { "login": "jbalam-nv", "id": 4916480, "node_id": "MDQ6VXNlcjQ5MTY0ODA=", "avatar_url": "https://avatars.githubusercontent.com/u/4916480?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jbalam-nv", "html_url": "https://github.com/jbalam-nv", "followers_url": "https://api.github.com/users/jbalam-nv/followers", "following_url": "https://api.github.com/users/jbalam-nv/following{/other_user}", "gists_url": "https://api.github.com/users/jbalam-nv/gists{/gist_id}", "starred_url": "https://api.github.com/users/jbalam-nv/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jbalam-nv/subscriptions", "organizations_url": "https://api.github.com/users/jbalam-nv/orgs", "repos_url": "https://api.github.com/users/jbalam-nv/repos", "events_url": "https://api.github.com/users/jbalam-nv/events{/privacy}", "received_events_url": "https://api.github.com/users/jbalam-nv/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "@lodm94 You are right, there seems to be an issue with the default value of noise_tar_paths. Until we push a fix, as a workaround you could change your dict to the following:\r\n\r\naudio_augmentations = dict(\r\n rir_noise_aug = dict(\r\n prob=0.5,\r\n rir_manifest_path = 'rir_manifest.json',\r\n rir_prob = 0.5,\r\n noise_manifest_paths = ['noise_manifest.json'],\r\n min_snr_db = [0],\r\n max_snr_db = [30],\r\n bg_noise_manifest_paths = ['noise_manifest.json'],\r\n bg_min_snr_db = [10],\r\n bg_max_snr_db = [40],\r\n noise_tar_filepaths=[None], \r\n bg_noise_tar_filepaths=[None]\r\n )\r\n)\r\n", "Hi @jbalam-nv thanks for the feedback. That was exactly how i managed to un the augmentation, i just forgot to reply here but it works.\r\nAnyway things went somehow bad. Take a look here:\r\n\r\n\r\n\r\nOrange line: model from pre-trained fine tuned with spec-augment only and 90h of data.\r\nBlue line: model from pre-trained fine tuned with spec-augment, rir_and_noise and 240h of data.\r\nIs there some tips you could give me to improve wer? Best result i got was 15.04% (ORAGE LINE)\r\nIf details are needed i will supply asap.\r\nThanks", "@lodm94 I would suggest reducing augmentation probability (0.1 or 0.2) . What is the learning rate for your fine-tuning?", "@jbalam-nv tomorrow i will run an experiment with reduced probs (either noise and rir). Here's my optim dict:\r\n\r\n```\r\noptim:\r\n name: novograd\r\n # _target_: nemo.core.optim.optimizers.Novograd\r\n lr: .001\r\n betas: [0.9, 0.25]\r\n weight_decay: 0.001\r\n\r\n sched:\r\n name: CosineAnnealing\r\n warmup_steps: null\r\n warmup_ratio: 0.12\r\n min_lr: 0.0\r\n last_epoch: -1\r\n```", "Sorry i accidentally closed the issue. One more question about optim: should i try a different min_lr? Should training stop when i reaach the min_lr or maybe it could be worth to let the model train for some epochs at costant min_lr?", "Hi @jbalam-nv i tried with reduced probs as you said. Training was going well but then\r\n\r\n\r\n\r\n\r\nduring validation check at the end of the epoch, i can see only the reference sentence. Decode is blank.\r\nDuring training, instead, i can see the decoded sentence and it is perfect, so perfect that i am afraid model is overfitting. Look that u-shape in val_loss. ", "Hi All, Thanks for this thread @titu1994 and @jbalam-nv. I followed these instructions and could successfully train a musan noise augmented nemo ASR model. However, I have few queries:\r\n\r\n1. Where does the noise augmented files are stored? Or Are these noise augmented files are simply produced on the fly, as and when needed, and discarded once the job is done (without retaining them)?\r\n\r\n2. When I use augmentors during training, does it include original clean files also for training along-with noise augmented files? Or it only uses noise augmented files for training?\r\n\r\n3. For the kind of experiments & datasets that we used, we had expected lower WERs with noise augmentation compared to WERs obtained on clean datasets. However the obtained results are other way (noise augmented datasets have 1% higher WER than clean data). This made us to suspect the implementation part of nemo asr training. I have used exactly the same config of **audio_augmentations** mentioned by @jbalam-nv. There were no errors during training as well. Do you still feel there might be some implementation gaps (which has led to lower WER). Any suggestions on this would be very helpful.\r\n\r\n4. I have used musan free sound noise and RIR. Is it fine? or any suggestions on this?\r\n\r\nThanks\r\nManjunath", "\r\nHi @titu1994 or @jbalam-nv , Could you please suggest me on my above query. Thanks", "Do not reopen discussing on a months old thread. Create a new one. " ]

[]

[]

[ "Hmm wait, the function exists in main but the error exists in b3? If the error exists only in main, then for now comment out these patches with a note that it needs to be uncommented. Or use a try catch. \r\n\r\nAlso, tests are failing on a notebook only pr?", "> Hmm wait, the function exists in main but the error exists in b3? If the error exists only in main, then for now comment out these patches with a note that it needs to be uncommented. Or use a try catch.\r\n> \r\n> Also, tests are failing on a notebook only pr?\r\n\r\n@titu1994 the error exists in b3 well the patch function exists in main. Is it safe to change branch in some notebooks to main now? If it's not, I can revert it back to b3 and fix it by manually accessing params or writing the patch function explicitly in notebooks. \r\n\r\nRight not sure why the tests fails for notebook changes ", "We will be changing branch later, main is not suitable for the notebooks cause it's not stable and tested as much as the exact branches.\n\nFor now, It might be best to do manual fixes by back porting the config updater to the notebooks.", "> We will be changing branch later, main is not suitable for the notebooks cause it's not stable and tested as much as the exact branches.\r\n> \r\n> For now, It might be best to do manual fixes by back porting the config updater to the notebooks.\r\n\r\nSounds great. will update it." ]

[]

[]

[ ">>\"Can I integrate it with my college website\"\r\n- Yes.\r\n\r\n>> \"could you please guide me?\"\r\n- Have a look at this notebook for example. \r\nhttps://colab.research.google.com/github/NVIDIA/NeMo/blob/r1.0.0b3/tutorials/NeMo_voice_swap_app.ipynb" ]

[]

[]

[ "This pull request **fixes 1 alert** when merging bdac5e9dff261cc3eb92ec5ce82746e0dab2a369 into a42987279f0226aea7d776840ff06a56c4bfb259 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-7fa88b8e67e895c005a893c335113b38bc7711b1)\n\n**fixed alerts:**\n\n* 1 for Except block handles 'BaseException'" ]

[]

[]

[]

[]

[]

[ "This pull request **introduces 1 alert** when merging be2e4ab037277ddc4832ac2dc36155fc3ddccc63 into db60a786c71767478a49dfcad772718f9d4878d2 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-97c11b343a84e9d6d8a330403161e07698220e3c)\n\n**new alerts:**\n\n* 1 for Except block handles 'BaseException'" ]

[]

[]

[ "currently we don't support such feature. Also, it would put high requirements on available RAM" ]

[]

[]

[ "Could you try by replacing gpus=[1] with gpus=1? This seems to be trying to index GPU [1] which would be the second gpu, but colab has just 1 gpu", "Hi Good Sir,\n\nYes, it seems that this is the case. It runs smoothly now once I have\nchanged to \"gpus=1\".\n\nAlthough would just like to inform you that looks like the an4 server the\nnotebook is fetching data from is currently down. I have to use my own data\nto validify your suggested solution.\n\nEither way, thank you for your kind help.\n\nRegards\n\nOn Thu, 21 Jan 2021 at 12:00, Somshubra Majumdar <[email protected]>\nwrote:\n\n> Could you try by replacing gpus=[1] with gpus=1? This seems to be trying\n> to index GPU [1] which would be the second gpu, but colab has just 1 gpu\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/NVIDIA/NeMo/issues/1653#issuecomment-764221827>, or\n> unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ASQ73YTFZPOJPPHYQF6RTYDS26REPANCNFSM4WMBQN4A>\n> .\n>\n" ]

[]

[]

[ "This pull request **introduces 2 alerts** when merging 1548c44f5dc90b95d7fc3d8fd1f1e83347753c74 into eb927fe6aa69f8d73705dcb785a2afb95e4c279d - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-2b1f14de5396282d23465f680fcf532753f62246)\n\n**new alerts:**\n\n* 1 for Non\\-callable called\n* 1 for Unused import" ]

[]

[]

[]

[]

[]

[]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[ { "login": "fayejf", "id": 36722593, "node_id": "MDQ6VXNlcjM2NzIyNTkz", "avatar_url": "https://avatars.githubusercontent.com/u/36722593?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fayejf", "html_url": "https://github.com/fayejf", "followers_url": "https://api.github.com/users/fayejf/followers", "following_url": "https://api.github.com/users/fayejf/following{/other_user}", "gists_url": "https://api.github.com/users/fayejf/gists{/gist_id}", "starred_url": "https://api.github.com/users/fayejf/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fayejf/subscriptions", "organizations_url": "https://api.github.com/users/fayejf/orgs", "repos_url": "https://api.github.com/users/fayejf/repos", "events_url": "https://api.github.com/users/fayejf/events{/privacy}", "received_events_url": "https://api.github.com/users/fayejf/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "Thanks for reporting this. We will fix it asap and let you know.", "@shfaizan [Fixed](https://github.com/NVIDIA/NeMo/pull/1659)! Please pull and update. Thanks!" ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[ { "login": "jbalam-nv", "id": 4916480, "node_id": "MDQ6VXNlcjQ5MTY0ODA=", "avatar_url": "https://avatars.githubusercontent.com/u/4916480?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jbalam-nv", "html_url": "https://github.com/jbalam-nv", "followers_url": "https://api.github.com/users/jbalam-nv/followers", "following_url": "https://api.github.com/users/jbalam-nv/following{/other_user}", "gists_url": "https://api.github.com/users/jbalam-nv/gists{/gist_id}", "starred_url": "https://api.github.com/users/jbalam-nv/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jbalam-nv/subscriptions", "organizations_url": "https://api.github.com/users/jbalam-nv/orgs", "repos_url": "https://api.github.com/users/jbalam-nv/repos", "events_url": "https://api.github.com/users/jbalam-nv/events{/privacy}", "received_events_url": "https://api.github.com/users/jbalam-nv/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "@lodm94 The issue here is that the script get_openslr_rir.py only works with absolute paths. \r\nPlease change the first line in the cell to:\r\n`rir_data_path = os.getcwd()`\r\nI will fix this in the main branch too. Thanks for letting us know of this issue.", "Holy moly it finally works! Cheers man! Thank you\r\nWhen fixing the main branch, consider that\r\n`!python ../../scripts/get_openslr_rir.py --data_root {rir_data_path}`\r\nreturns error. I had to wget the script and change this path.\r\n\r\n" ]

[]

[]

[ "The type error mentions that \"input_signal_length\" has no corresponding input types - and the two valid input names are input_signal and length. \r\n\r\nSo if you replace the call function argument of empty_model.preprocessor() from input_signal_length to just length, it should work ", "Thank you, I was able to get this working thanks to your help and will share code in case anyone else needs to replicate. That being said I'm still hoping there is a better way to do this. I also had to change normalization in config from `per_feature` to `all_features` to avoid a NaN issue. \r\n\r\nAny insight into why there are two channels returned by the preprocessor and one of them is empty? Is it because I pass the length as a 2D tensor (shape: [1, num_samples]) instead of a 1D tensor? Thank you.\r\n\r\n```\r\nfrom IPython.display import display\r\nimport matplotlib.pyplot as plt\r\nimport librosa.display\r\nplt.rcParams[\"figure.figsize\"]=(12,9)\r\ndef display_specs(model, audio_file):\r\n display(Audio(audio_file))\r\n model.to('cpu')\r\n y0,sr = torchaudio.load(audio_file)\r\n y0r = torchaudio.transforms.Resample(sr,16000)(y0)\r\n y0_len = torch.tensor(y0r.shape)\r\n fig, ax = plt.subplots(nrows=3, ncols=1, sharex=True)\r\n spec_result = model.preprocessor(input_signal=y0r, length=torch.tensor(y0_len))\r\n ax[0].set(title=\"First channel of spec\")\r\n librosa.display.specshow(spec_result[0][0].numpy(), ax=ax[0])\r\n ax[1].set(title=\"Second channel of spec\")\r\n librosa.display.specshow(spec_result[0][1].numpy(), ax=ax[1])\r\n aug_result = model.spec_augmentation(input_spec=spec_result[0][1].unsqueeze(0))\r\n ax[2].set(title=\"Augmented second channel of spec\")\r\n librosa.display.specshow(aug_result[0].numpy(), ax=ax[2])\r\n```\r\n\r\nUse by calling \r\n`display_specs(<your_model_instance>, <path_to_audiofile>)`", "> I also had to change normalization in config from per_feature to all_features to avoid a NaN issue.\r\n\r\nThis should not be needed\r\n\r\n> Any insight into why there are two channels returned by the preprocessor and one of them is empty? Is it because I pass the length as a 2D tensor (shape: [1, num_samples]) instead of a 1D tensor? \r\n\r\nLength is supposed to be a 1D tensor with each element representing the duration per sample. So len(length) == batch size; length[0] = len(sample_0).\r\n\r\nYou can take a look at `FilterbankFeatures` to make out whats happening inside the preprocessor module. \r\n" ]

[]

[]

[ "Hi! Thanks for the PR. Could you sign your commit using the instructions here - https://github.com/NVIDIA/NeMo/pull/1644/checks?check_run_id=1721045606", "@titu1994 I cannot read the details of check, please check it. Thanks" ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[]

[ "You don't need the decoder.params.vocabulary anymore. Directly use decoder.vocabulary", "Yes this solves the issue thanks" ]

[]

[]

[]

[]

[ { "login": "ekmb", "id": 10428420, "node_id": "MDQ6VXNlcjEwNDI4NDIw", "avatar_url": "https://avatars.githubusercontent.com/u/10428420?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ekmb", "html_url": "https://github.com/ekmb", "followers_url": "https://api.github.com/users/ekmb/followers", "following_url": "https://api.github.com/users/ekmb/following{/other_user}", "gists_url": "https://api.github.com/users/ekmb/gists{/gist_id}", "starred_url": "https://api.github.com/users/ekmb/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ekmb/subscriptions", "organizations_url": "https://api.github.com/users/ekmb/orgs", "repos_url": "https://api.github.com/users/ekmb/repos", "events_url": "https://api.github.com/users/ekmb/events{/privacy}", "received_events_url": "https://api.github.com/users/ekmb/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "I am also encountering this problem. Tried setting num_workers=0 in _setup_infer_dataloader method of nemo/collections/nlp/models/token_classification/punctuation_capitalization_model.py, but it did not solve problem.", "We don’t support Windows at this point, please try running in Colab `https://colab.research.google.com/github/NVIDIA/NeMo/blob/r1.0.0rc1/tutorials/NeMo_voice_swap_app.ipynb`", "@ekmb I don't know if you're still interested, but I actually ended up fixing this problem by changing `DEFAULT_PROTOCOL` in torch/serialization.py to 4 followed by changing all `num_workers` to 0 in punctuation_capitalization_model.py" ]

[ { "id": 1719393562, "node_id": "MDU6TGFiZWwxNzE5MzkzNTYy", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/feature", "name": "feature", "color": "68F3F0", "default": false, "description": "request/PR for a new feature" } ]

[ { "login": "okuchaiev", "id": 22035961, "node_id": "MDQ6VXNlcjIyMDM1OTYx", "avatar_url": "https://avatars.githubusercontent.com/u/22035961?v=4", "gravatar_id": "", "url": "https://api.github.com/users/okuchaiev", "html_url": "https://github.com/okuchaiev", "followers_url": "https://api.github.com/users/okuchaiev/followers", "following_url": "https://api.github.com/users/okuchaiev/following{/other_user}", "gists_url": "https://api.github.com/users/okuchaiev/gists{/gist_id}", "starred_url": "https://api.github.com/users/okuchaiev/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/okuchaiev/subscriptions", "organizations_url": "https://api.github.com/users/okuchaiev/orgs", "repos_url": "https://api.github.com/users/okuchaiev/repos", "events_url": "https://api.github.com/users/okuchaiev/events{/privacy}", "received_events_url": "https://api.github.com/users/okuchaiev/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "Adding a `me too` comment here. I'm actually interested in loading features archives computed by kaldi (like fbank). This is a feature very useful we find in other toolkits like ESPnet for example. Is there some documentation or some guidelines how one would implement that?", "@jprobichaud I already fixed here\r\nhttps://github.com/NVIDIA/NeMo/pull/1644", "@qmpzzpmq Thanks for your reply. This is useful for loading audio that can come from transformation, but that assumes that the entire audio will be one \"segment\". The kaldi feature data layer allowed if I read properly to read in extracted features for each segments (one audio file could be split in multiple shorter segments.\r\n\r\nIt's already nice to have the pipe reading available. Thanks for that! I still wonder if reading audio + segment or raw features will be possible again in the future or if that was deem not useful.\r\n", "Hi @jprobichaud and @qmpzzpmq Thanks for your requests and contributions.\r\n\r\nYes we think the external features (like fbank) is quite useful. Due to some maintenance reason, we don't have KaldiFeatureLayer now. But we are working on ExternalFeatureDataloader that reads the stored feature (in pickle) and we will provide the script to extract and store it to pickle files. Hope this would give more flexibility. We might support reading from feature.ark as well later. \r\n\r\n> extracted features for each segments (one audio file could be split in multiple shorter segments.\r\n\r\nLike [feat1, feat2, feat3] corresponding to [utt1, utt2, utt3]? @jprobichaud could you elaboration more so we can understand your idea clearly? ", "@jprobichaud \r\nusing \"utils/data/extract_wav_segments_data_dir.sh\"" ]

[]

[]

[]

[]

[]

[ "While not consistent with the paper, this is a tutorial to explain how to refactor a monolithic concrete model into an abstract NeMo model - comprising of neural modules that represent the abstraction with neural types, and then instantiate them with actual concrete modules (which can be anything that fits the neural types) as a Nemo Model. \r\n\r\nThe concept of a NeMo models components don't need to be 1:1 with a papers notation of terms - they are represented as \"encoder\" that takes certain neural type and returns some neural type, and a \"decoder\" that takes some neural type from the encoder and returns some other neural type. They don't refer to the encoder and decoder blocks of the gpt blocks only. \r\n\r\nSee ASR models who also present encoder decoder architecture. \r\n\r\nThe encoder neural module can be a cnn based QuartzNet or a self attention based transformer decoder block or a Conformer sandwich. \r\n\r\nSimilarly, the decoder neural module can be a basic linear layer, a RNN based Autoregressive decoder, or even an RNNT Transducer based decoding scheme. \r\n\r\nFor now, to be clear I will add a note there" ]

[]

[]

[]

[]

[]

[ "Hmmm, I think the line that was breaking NeMo installation on my MacBook was in `setup.py`:\r\nhttps://github.com/NVIDIA/NeMo/blob/45f60f03e5d15a4b6e7db145b5d731d73f3a891d/setup.py#L235\r\n\r\n@borisfom ideas?", "> Hmmm, I think the line that was breaking NeMo installation on my MacBook was in `setup.py`:\r\n> https://github.com/NVIDIA/NeMo/blob/45f60f03e5d15a4b6e7db145b5d731d73f3a891d/setup.py#L235\r\n> \r\n> @borisfom ideas?\r\n\r\n@tkornuta-nvidia Doesn't seem like this PR breaks MacOS compatibility. It looks like it was broken before? This PR change reduces ```setup_requires=['pytest-runner', 'nvidia-pyindex'],``` to ```setup_requires=['pytest-runner']```", "Awaiting @borisfom review before merge", "Merged as main branch is broken without this, preventing other prs ", "Sorry to be late for the game - I am on vacation right now ...\r\nToo bad onnx-graphsurgeon is not available on all platforms - we would need to think how to handle those cases properly in general. It makes it rather difficult to have seamless setup.py, without hairy scripts around. As my original change was primarily to enable a workaround for TRT issue that will hopefiully be fixed, it's probably fine in this case. \r\nThe way it's left after this PR merged is not right though: \r\n1. constant_folding argument is defunct (always overridden). We should revert changes to it and only have onnx-graphsurgeon invocation under the GRAPGSURGEION_AVAILABLE conditional.\r\n2. default setup.py would not install onnx-graphsurgeon on any platform. That's probably the best we can do at the moment and have developers who need to use this WAR, to install it manually - but we should think of general way of handling optional components to have best experience on all platforms. It sees that, besides WAR for TRT 'double' issue, onnx-graphsurgeon pass makes most of the graphs more compact (potentially faster, one needs to check), so it would probably be nice to keep it enabled for the platforms that have it, even when TRT issue is fixed." ]

[]

[]

[]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[]

[ "This is a good find, I keep forgetting TF and PT have different semantics for their RNNs. \r\n\r\nNeMo currently ignores nested structures (h, c) (though we plan to add this), so that will be a todo for a subsequent patch. In the diagram above, \"h\" corresponds to \"y_t\" and \"c\" corresponds to both (h_n, c_n), so a bit of renaming will be necessary as well.\r\n\r\nAs to the axes, we cant detach the num_directions * num_layers since it flattens out into a single axis. The axiskind most appropriate to that would probably be Channel/Dimension rather than Any. Any is a bit too generic for an RNN state axis.\r\n\r\n" ]

[]

[ { "login": "fayejf", "id": 36722593, "node_id": "MDQ6VXNlcjM2NzIyNTkz", "avatar_url": "https://avatars.githubusercontent.com/u/36722593?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fayejf", "html_url": "https://github.com/fayejf", "followers_url": "https://api.github.com/users/fayejf/followers", "following_url": "https://api.github.com/users/fayejf/following{/other_user}", "gists_url": "https://api.github.com/users/fayejf/gists{/gist_id}", "starred_url": "https://api.github.com/users/fayejf/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fayejf/subscriptions", "organizations_url": "https://api.github.com/users/fayejf/orgs", "repos_url": "https://api.github.com/users/fayejf/repos", "events_url": "https://api.github.com/users/fayejf/events{/privacy}", "received_events_url": "https://api.github.com/users/fayejf/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "> After training the Voice Activity Detection model, How to deploy it for inferencing on arm based devices such as Raspberry Pi?\r\n\r\nWe dont have deployment scripts for Raspberry Pi. \r\n\r\n> How much memory is required to inference MatchboxNet\r\n\r\nDepends on deployment framework, but it should be small considering the size of the model. \r\n\r\n> How does MatchboxNet perform against MarbleNet\r\n\r\nThey perform two different (but similar tasks) so they cant be compared directly. Matchboxnet is meant for speech command recognition - multiclass classification problem and Marblenet is meant for voice activity detection - binary classification problem. ", "- For the Voice Activity Detection task, the deployment to ARM-based devices is vital and should be considered.\r\nWhat difficulties need to be solved to be able to deploy the VAD model for inferencing on an Raspberry Pi?\r\n- isn't Matchboxnet currently used for VAD? " ]

[]

[]

[ "Hi @skubicius Thanks for your question 😃 . This is the expected plot for this demo wav. There are a few reasons for a not ideal plot for the pre-trained models, such as data and sample rate mismatch, diverse recording (microphone) environment, language, etc. And please note that metrics would be different for different use cases, for example, should we treat short pause as no-speech? It depends. We would encourage you to further tune the model (such as model finetuning, threshold optimization) on your training/development data to boost the performance on unseen data. In the meanwhile, we are trying to release better pre-trained models and will update the notebook with the necessary clarification and guide. ", " Hi @fayejf, thank you for the explanation, I'll check out model tuning.", "No prob! @skubicius You should be able to find instructions in [tutorial 06 ](https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/06_Voice_Activiy_Detection.ipynb). Please let us know if you have more questions.", "hi @fayejf \r\nI trained VAD model using my own dataset follow instructions in [tutorial 6](https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/06_Voice_Activiy_Detection.ipynb)\r\nMy dataset is Vietnamese telephony conversation, 15 hours with 8 hours of speech and 7 hours of noise. I got accuracy 95% on my test set.\r\nThe format of dataset quite similar with google speech dataset. Each sample of speech is one statement and have duration 2s to 5s.\r\n\r\nWhen i follow instruction in [tutorial 7](https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/07_Online_Offline_Microphone_VAD_Demo.ipynb) to inference sample in test set what contain only one statement, the prediction is good. The problem is when i inference long audio has duration 20s of speech and noise, the prediction is really bad.\r\nThis is predict sample of speech has only one statement \r\n\r\n\r\n\r\nAnd long audio, 0 to 5s of audio is noise.\r\n\r\n\r\n\r\n1. Any idea to deal with this problem\r\n2. I want to tune threshold but it required RTTM ground truth, what is the format of RTTM file and how to generate this file from dataset like google speech.\r\n", "@guyqaz \r\n\r\nThanks for your question. \r\nWhat is the sample rate of your long audio sample? Looks like it misses some predictions after 15s.\r\n\r\nYou could find the definition of NIST Rich Transcription Time Marked (RTTM) format [here](https://web.archive.org/web/20170119114252/http://www.itl.nist.gov/iad/mig/tests/rt/2009/docs/rt09-meeting-eval-plan-v2.pdf). We have prepared an example file for you [here](https://nemo-public.s3.us-east-2.amazonaws.com/an4_diarize_test.rttm). For google speech basically you can have something like.\r\n\r\n`SPEAKER File1 1 0.5 0.671 <NA> <NA> Speaker1 <NA> <NA>`\r\n`SPEAKER File1 1 1.7 1.258 <NA> <NA> Speaker1 <NA> <NA>`\r\n\r\nTo tune threshold, the RTTM file **is not required**. If you have frame level label, for example, [1,1,1,1,1,0,0,0,0,1,1,1], you can directly use that as well. \r\n\r\nNote the flow is \r\nframe level prediction (w/ or wo/posterior) -> to convert to speech / no-speech given **threshold** -> compare with ground truth label to find optimal threshold\r\n", "Hi @fayejf , The sample rate of long audio sample is 16000Hz.\r\nNow i am defining one audio of sentence as speech, not a word level like google speech dataset. But i still using [process_vad_data](https://github.com/NVIDIA/NeMo/blob/main/scripts/dataset_processing/process_vad_data.py) to generate manifest file.\r\nWhen infer audio just has statement or noise, i think it good because almost all frame are predicted to be speech or all noise. But when predict long audio have both speech and noise. Model dose not seen recognize speech and noise.\r\n\r\nMy dataset also have label with shorter duration what label for stop point, noise in audio contain speech of statement. \r\n- Do you think i should using shorter label or just using audio of sentence is speech.\r\n- Is there a tutorial to tune threshold from output of predict frame level [vad_infer](https://github.com/NVIDIA/NeMo/blob/main/examples/asr/vad_infer.py) and manifest file.", "@guyqaz \r\nAnswered here https://github.com/NVIDIA/NeMo/issues/2353#issuecomment-861862612" ]

[]

[]

[]

[]

[ { "login": "blisc", "id": 4763352, "node_id": "MDQ6VXNlcjQ3NjMzNTI=", "avatar_url": "https://avatars.githubusercontent.com/u/4763352?v=4", "gravatar_id": "", "url": "https://api.github.com/users/blisc", "html_url": "https://github.com/blisc", "followers_url": "https://api.github.com/users/blisc/followers", "following_url": "https://api.github.com/users/blisc/following{/other_user}", "gists_url": "https://api.github.com/users/blisc/gists{/gist_id}", "starred_url": "https://api.github.com/users/blisc/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/blisc/subscriptions", "organizations_url": "https://api.github.com/users/blisc/orgs", "repos_url": "https://api.github.com/users/blisc/repos", "events_url": "https://api.github.com/users/blisc/events{/privacy}", "received_events_url": "https://api.github.com/users/blisc/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[]

[]

[]

[ "This pull request **introduces 1 alert** when merging 657e3e50296a8181164b830a812b40858d5917a6 into bc5f03459c7273b71c50af7c2378d5e950b408c4 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-a3b2ef6ca144959b67e47fc21de10d8f7c6251b2)\n\n**new alerts:**\n\n* 1 for &#39;import \\*&#39; may pollute namespace", "Hi @titu1994, I would like to merge this pull request. Could you review it? Thanks in advance", "This pull request **introduces 1 alert** when merging b4cbe146d2d376a805cdb183615068a06d38683a into bc5f03459c7273b71c50af7c2378d5e950b408c4 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-8ac0d3b412e9f286a1c6b459f6d1ec28aaf35deb)\n\n**new alerts:**\n\n* 1 for &#39;import \\*&#39; may pollute namespace", "I completely agree with your comments, there are clear opportunities to avoid code duplication. I've implemented this way for the sake of clarity and to avoid conflicts. Anyway, I will watch the related PR and refactor this one after it's been merged. \r\n\r\nOn the other hand, the DCO check is failing as my commits have been signed with a wrong email. I've tried to correct it by rebasing, but I wasn't able to fix it so far. Any clue about how to handle it?\r\n\r\nThanks for your help", "HI @diego-fustes Sorry for keep you waiting. [PR#1617](https://github.com/NVIDIA/NeMo/pull/1617) is merged now 😃 . Please rebase on it. And let me know if you have any questions.", "This pull request **introduces 1 alert** when merging 33d6fd0525ba7728f9d34aaa19fa45f29d038c47 into 14904fd0381f5515a9773cf7a538dc09e08c8aa9 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-ff942955051d87be1008417b4aab787ce24f8579)\n\n**new alerts:**\n\n* 1 for &#39;import \\*&#39; may pollute namespace", "Hi @fayejf and @titu1994 :\r\n\r\nI've been looking at the current code base, and I think that some refactoring is still needed. To start with, I think that EndDecSpeakerLabelModel should be a subclass of EncDecClassificationModel. At the end, I think that it's almost the same, just with the option of switching the loss to AngularSoftmaxLoss and that it outputs the embeddings along with the logits. \r\n\r\nOn the other hand, the class EncDecRegressionModel that I'm implementing should be a superclass of EncDecClassificationModel. At the end, classification can be seem as a special type of regression with a threshold...\r\n\r\nWhat do you think? It's quite a major change, I'm willing to implement it if you agree ", "Hey @diego-fustes,\r\n\r\nYou raise some are valid points, but there are other reasons they are kept separate as of now.\r\n\r\n> EndDecSpeakerLabelModel should be a subclass of EncDecClassificationModel. At the end, I think that it's almost the same, just with the option of switching the loss to AngularSoftmaxLoss and that it outputs the embeddings along with the logits.\r\n\r\nYes, this is true - for the present. The reason we internally decided to separate the two models is that Speech classification is guarenteed to have a 1 ip - 1 op pipeline as in general classification models. Speaker classification on the other hand may use auxiliary information in the future, or might require architectures with multiple outputs. Speaker classification is also one step in the entire speaker diarization pipeline, which needs different interaction than the simpler classification model. All in all, it was therefore concluded that albeit they share significant similarities, they would still be separate entities.\r\n\r\n> EncDecRegressionModel that I'm implementing should be a superclass of EncDecClassificationModel. At the end, classification can be seem as a special type of regression with a threshold\r\n\r\nThis is something I disagree with. Simply put, while mathematically correct that classification is a special case of regression, that does not mean it is semantically the same task. We dont apply regression losses (MSE) to train MNIST / ImageNet, nor do we use any activation function for regression models (as an example). The dataset, and the semantic task that the model performs is entirely different, even if mathematically the only difference is the application of softmax/sigmoid at the end of the model.\r\n", "This pull request **introduces 1 alert** when merging 01e79acda2faf0f9b38ef58d73bc1564a10cb4ae into 941ef1fd71bd2515a4ba7092d65146edfddc1229 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-460d53983a0a2e4ee69ca4a17081e35c2bc9dda9)\n\n**new alerts:**\n\n* 1 for &#39;import \\*&#39; may pollute namespace", "Hi @titu1994 \r\n\r\nI agree with you, from a practical perspective it's better to implement the common code for regression and classification in an abstract class. I've just pushed a new version, trying to minimize code duplication. Apologies as I think that I've added more commits that needed, I'm a kind of starting with GitHub.\r\n\r\nKind regards", "This pull request **introduces 1 alert** when merging 1452075151885cb3a3395d768d4603dac6c16ef5 into 941ef1fd71bd2515a4ba7092d65146edfddc1229 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-8cc3f74e16c23909b4cedfe02da85605b1763535)\n\n**new alerts:**\n\n* 1 for &#39;import \\*&#39; may pollute namespace", "Hey @diego-fustes, no worries about Git. Could you try rebasing your commit on top of main branch instead of merge? That should clean up the commit history. \r\n\r\nI was looking though some of the changes, and they are pretty good. Some minor cleanup here and there and it should be good to merge. I'll post the comments after the rebase (as comments sometime get deleted after rebase)" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Overall, the PR looks good to merge. Please fix the tests and final review from @MaximumEntropy please. " ]

[]

[]

[]

[]

[ { "login": "nithinraok", "id": 19668129, "node_id": "MDQ6VXNlcjE5NjY4MTI5", "avatar_url": "https://avatars.githubusercontent.com/u/19668129?v=4", "gravatar_id": "", "url": "https://api.github.com/users/nithinraok", "html_url": "https://github.com/nithinraok", "followers_url": "https://api.github.com/users/nithinraok/followers", "following_url": "https://api.github.com/users/nithinraok/following{/other_user}", "gists_url": "https://api.github.com/users/nithinraok/gists{/gist_id}", "starred_url": "https://api.github.com/users/nithinraok/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/nithinraok/subscriptions", "organizations_url": "https://api.github.com/users/nithinraok/orgs", "repos_url": "https://api.github.com/users/nithinraok/repos", "events_url": "https://api.github.com/users/nithinraok/events{/privacy}", "received_events_url": "https://api.github.com/users/nithinraok/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "It is implemented for a WIP project now, It could be done using collate func to slice the batches and combine back them by backtracking the values. \r\n\r\nFor reference see here:\r\nhttps://github.com/NVIDIA/NeMo/blob/df360cd3c856c645d23505305ee8b689636b2dd1/nemo/collections/asr/data/audio_to_label.py#L161\r\n\r\nAnd we capture outputs here:\r\nhttps://github.com/NVIDIA/NeMo/blob/df360cd3c856c645d23505305ee8b689636b2dd1/nemo/collections/asr/models/label_models.py#L407 " ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[]

[ "@rahul-art should be fixed now" ]