Datasets:

---

renwei2024

/

nemo-github-issues

like 0

[]

[]

[]

[]

[]

[ "Seems theres some formatting errors, could you run `python setup.py style --scope nemo --fix`", "This PR will be merged into another PR (https://github.com/NVIDIA/NeMo/pull/2240) due to inactivity. One difference is that it will not raise the error, but instead simply log that the file was not loaded properly.", "Got it, sorry for the delay and thanks for the followup!", "Thanks for the PR! Normally I'd wait, but this PR would have merge conflicts with that one so it would be easier to integrate into that one. " ]

[]

[]

[ "We generally don't prefer to have other file formats for efficiency reasons (wav takes storage but at least it prevents on the fly resampling to 16khz monochannel audio). \r\n\r\nTarred datasets are meant for distributed training on clusters, where network I/O can have a significant impact on training speed. Likewise, it would be unwise to spend compute for such expensive resampling on the CPU repeatedly per sample.\r\n\r\nIn any case, I'll look into what's currently preventing Tarred datasets of ASR from working with other file formats. ", "I should've left a more detailed trace, sorry.\r\nThe issue happens [here](https://github.com/NVIDIA/NeMo/blob/489bb241a6ed727b4d7ad3213242772a7628ff55/nemo/collections/asr/data/audio_to_text.py#L920) where the `webdataset` try's to rename the audio key from `wav`.\r\n\r\nInstead, this should be an arg passed to `_TarredAudioToTextDataset`.\r\n\r\nOtherwise, you get the following trace \r\n\r\n```\r\nOriginal Traceback (most recent call last): [1447/1909]\r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py\", line 198, in _worker_loop \r\n data = fetcher.fetch(index) \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py\", line 28, in fetch \r\n data.append(next(self.dataset_iter)) \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/webdataset/iterators.py\", line 226, in map \r\n for sample in data: \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/nemo/collections/asr/data/audio_to_text.py\", line 754, in __next__ \r\n audio_bytes, audio_filename = next(self.iterator) \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/webdataset/iterators.py\", line 282, in to_tuple \r\n for sample in data: \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/webdataset/iterators.py\", line 244, in rename \r\n if handler(exn): \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/webdataset/utils.py\", line 12, in reraise_exception \r\n raise exn \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/webdataset/iterators.py\", line 242, in rename \r\n yield {k: getfirst(sample, v, missing_is_error=True) for k, v in kw.items()} \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/webdataset/iterators.py\", line 242, in <dictcomp> \r\n yield {k: getfirst(sample, v, missing_is_error=True) for k, v in kw.items()} \r\n File \"/home/ubuntu/.local/lib/python3.8/site-packages/webdataset/iterators.py\", line 77, in getfirst \r\n raise ValueError(f\"didn't find {keys} in {list(a.keys())}\") \r\nValueError: didn't find ['wav'] in ['__key__', 'mp3'] \r\n``` ", "That makes sense, thanks for the hint! " ]

[]

[]

[ "These docs are quite outdated, could you point me to where they were referenced so we can update them ?\r\n\r\nThese are the latest docs - https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/\r\n\r\nThis section (https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html) has the steps to build KenLM and train a KenLM model, plus perform beam search using it. ", "Thanks! This answer has a 404 link, so I searched for the link I shared: https://github.com/NVIDIA/NeMo/issues/577\r\n\r\nIn the link above, I see that it mentions install_beamsearch_decoders.sh to install the decoders. But does this build the KenLM binary folder (kenlm/build/bin)? When I run install_beamsearch_decoders.sh, then run train_kenlm.py, it gives me the error decoders/kenlm/build/bin/lmplz is not found, so I believe I am missing a step where I build the binary?\r\n\r\nThanks!\r\n\r\nEDITED with fixed github issue link", "I fixed my user errors, but wanted to share what worked for me if others try to compile KenLM on Google Colab:\r\n\r\n1) I updated cmake to latest version: 3.20, using method B here: https://askubuntu.com/questions/355565/how-do-i-install-the-latest-version-of-cmake-from-the-command-line (takes about 10 minutes, and I hard-coded the version and build numbers, as it failed with variables for some reason)\r\n2) I ran KenLM build instructions after running install_beamsearch_decoders.sh, using here: https://github.com/kpu/kenlm/blob/master/BUILDING\r\n\r\nThank you so much, Somshubra, for pointing me in the right direction and responding so quickly! If I could make a constructive comment, I might add into the user guide you linked to that one needs to explictly compile KenLM after running install_beamsearch_decoders.sh. I had been unsure if that was being done by the script itself (since it does download KenLM and creates a build folder).", "That's a very good point, we'll update the docs to reflect that KenLM must still be compiled after performing the setup steps. @vsl9 @VahidooX fyi", "The script 'install_beamsearch_decoders.sh' has been already updated to install the binary kenlm. I think @vsl9 can verify this.\r\nHere is the commit in the openseq2seq:\r\nhttps://github.com/NVIDIA/OpenSeq2Seq/commit/5cb391bba14a09142b82c68c5de17d52ab98b51f\r\n\r\nDid you get any error message when ran the script?\r\n\r\nIf the script could not install the binary kenlm, then it can be a bug or some of the dependencies of the kenlm was not installed. You may find more info on its dependencies here:\r\nhttps://kheafield.com/code/kenlm/dependencies/\r\n\r\nI am going to add more hints and instructions for tackling such issues in the doc. ", "> The script 'install_beamsearch_decoders.sh' has been already updated to install the binary kenlm. I think @vsl9 can verify this.\r\n> Here is the commit in the openseq2seq:\r\n> [NVIDIA/OpenSeq2Seq@5cb391b](https://github.com/NVIDIA/OpenSeq2Seq/commit/5cb391bba14a09142b82c68c5de17d52ab98b51f)\r\n> \r\n> Did you get any error message when ran the script?\r\n> \r\n> If the script could not install the binary kenlm, then it can be a bug or some of the dependencies of the kenlm was not installed. You may find more info on its dependencies here:\r\n> https://kheafield.com/code/kenlm/dependencies/\r\n> \r\n> I am going to add more hints and instructions for tackling such issues in the doc.\r\n\r\nYes, at first the script could not install the binary kenlm because I was missing Boost package, so I had to install dependencies as you mentioned :\r\n`sudo apt-get install build-essential libboost-all-dev cmake zlib1g-dev libbz2-dev liblzma-dev` \r\nAfter that, I rerun `install_beamsearch_decoders.sh` script and everything were installed correctly. " ]

[]

[]

[ "Just a note - The manifest file can (and generally it should) have absolute paths to the dataset files when on a local machine, which the tarfile would then use. ", "> Just a note - The manifest file can (and generally it should) have absolute paths to the dataset files when on a local machine, which the tarfile would then use.\r\n\r\nYup, I noticed this. However, I was splitting based on specific audio/label heuristics unrelated to the train/val/test split.\r\nI'll probably modify my code to make reading these splits better." ]

[]

[ { "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 } ]

[ "Hi, again! I just tried with the `hi-mia` dataset and received the same error ", "Hi, I just ran `speaker_reco_finetune.py` and everything looks good to me. \r\nCan you make sure you have set the following params in conf/SpeakerNet_verification_3x2x512.yaml:\r\ntrain_set\r\ndev_set\r\ndecoder.emb_sizes: 256\r\ndecoder.num_classes: (num of uniq classes in your train set) \r\noptim.name: sgd\r\noptim.lr: 0.006 (# make sure to remove betas from args and replace with momentum) \r\nmax_epochs: 10 \r\n\r\nAnd use --pretrained_model=\"speakerverification_speakernet\" while running speaker_reco_finetune.py script(speakerrecognition also works, change pretrained_model name accordingly)", "Hi, @nithinraok Thank you for your reply! I made sure train_set, dev_set, and num_classes all correct, though the default optim.name was novograd and the max_epoch I went with was 5, I am changing them to the numbers you said. Though can I ask how to replace betas with momentum? I did not just find useful documentation for nemo configurations...Thank you again for your help! ", "btw I set gpus: 0, since torch.cuda.is_available() returns false (according the colab tutorial), could this be the cause of the problem? Does fine-tuning the pretrained model absolutely require gpu?", "```\r\n optim:\r\n name: sgd\r\n lr: .006\r\n # optimizer arguments\r\n args:\r\n name: auto\r\n weight_decay: 0.001\r\n momentum: 0.9 # replaced here\r\n```\r\n\r\nYes, I just tried with cpu, looks like optimizer running into an issue. Can you run on a gpu machine meanwhile, while I find out what could have been the cause of it. Thanks. ", "Thank you! I will try running on a GPU machine. First I will need to set up a few things and make sure torch.cuda.is_available() returns true. \r\nbtw I just changed the configuration file according to your descriptions with cpu and I get a different error this time\r\n```\r\nTraceback (most recent call last):\r\n File \"/Users/xujinghua/speaker-recognition-nemo/scripts/experimental/fine_tune_an4.py\", line 119, in <module>\r\n main()\r\n File \"/Users/xujinghua/speaker-recognition-nemo/scripts/experimental/fine_tune_an4.py\", line 76, in main\r\n finetune_config = OmegaConf.load(args.finetune_config_file)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/omegaconf/omegaconf.py\", line 254, in load\r\n obj = yaml.load(f, Loader=get_yaml_loader())\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/yaml/__init__.py\", line 114, in load\r\n return loader.get_single_data()\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/yaml/constructor.py\", line 43, in get_single_data\r\n return self.construct_document(node)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/yaml/constructor.py\", line 47, in construct_document\r\n data = self.construct_object(node)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/yaml/constructor.py\", line 92, in construct_object\r\n data = constructor(self, node)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/omegaconf/_utils.py\", line 95, in no_duplicates_constructor\r\n value = loader.construct_object(value_node, deep=deep)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/yaml/constructor.py\", line 92, in construct_object\r\n data = constructor(self, node)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/omegaconf/_utils.py\", line 95, in no_duplicates_constructor\r\n value = loader.construct_object(value_node, deep=deep)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/yaml/constructor.py\", line 92, in construct_object\r\n data = constructor(self, node)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/omegaconf/_utils.py\", line 95, in no_duplicates_constructor\r\n value = loader.construct_object(value_node, deep=deep)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/yaml/constructor.py\", line 92, in construct_object\r\n data = constructor(self, node)\r\n File \"/Users/xujinghua/miniconda3/lib/python3.7/site-packages/omegaconf/_utils.py\", line 101, in no_duplicates_constructor\r\n key_node.start_mark,\r\nyaml.constructor.ConstructorError: while constructing a mapping\r\n in \"/Users/xujinghua/speaker-recognition-nemo/conf/SpeakerNet_recognition_3x2x512.yaml\", line 126, column 7\r\nfound duplicate key name\r\n in \"/Users/xujinghua/speaker-recognition-nemo/conf/SpeakerNet_recognition_3x2x512.yaml\", line 132, column 7\r\n```\r\nThank you again for your help! I will first try gpu\r\n", "You are using a key name twice in config file. In the error you posted looks like the key is in line 126 and line 132. ", "ah, sorry! that was an oversight, it turned out to be an indentation error in the config file. I corrected the indentation and now it gets the same `ZeroDivisionError` again...", "Can you try running locally if possible? Probably something with colab", "I was running locally (still training on cpu though), still trying to set up gpu, then I will try again", "Hi @nithinraok I set up cuda, gpu and everything, this time instead of errors it shuts down my computer at the beginning of training...", "Not sure, why you are facing such issues. Let's take this step by step. Instead of finetuning are you able to train? \r\n", "I just tried it again and now it is working! I am not sure why the computer shut down in the previous tries, probably due to low battery... As a turns out the error were indeed caused by cpu training, I think to fine-tune the pretrained model it is mandatory to train on gpu too (which might be a small problem for mac users who don't have a NVIDIA graphics card :p). Anyways, thank you very much for your help & the great work of building and training the models!", "Hi @JINHXu ... To finetune on cpu, you have to make sure to set accelerator to `null` in config, along with gpus=0. This should work ", "Hi @nithinraok I just tried, it works for me too. Thank you again for all your help!" ]

[ { "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 } ]

[ "I couldn't reproduce any training issues with [this notebook](https://colab.research.google.com/github/NVIDIA/NeMo/blob/r1.0.0rc1/tutorials/nlp/Relation_Extraction-BioMegatron.ipynb) running on Colab. A few warnings do appear during installation but they can be disregarded. \r\nCould you please specify the exact cell you're having errors in? ", "I am also facing the same issue of No module named 'apex.normalization'.", " if you use NeMo container it has nvidia apex installed, otherwise you need use this link https://github.com/NVIDIA/apex/tree/master/apex " ]

[]

[]

[]

[]

[]

[ "which pip version do you have? Could be a problem with your environment. Are you able to pip install other packages? Like this: ```pip install pytorch-lightning==1.1.4``` ? I literally copy-pasted this: ```pip install nemo_toolkit[all]==1.0.0rc1``` and it worked." ]

[]

[]

[ "This pull request **introduces 1 alert** when merging 85b00da1038709cc1b145f28c8e8dd043efb2932 into d1b37ed30a70bff00df4778b55b020d8bab2e6fb - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-df3f0c730d38a856855a25d9a48542d309857460)\n\n**new alerts:**\n\n* 1 for Variable defined multiple times", "This pull request **introduces 1 alert** when merging 76792ea5da58ca76b2b0aea6690b314041d2334b into 040b21fd72618113a2987406a6a4bc0b910b2b3b - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-b9f1a282982ed548829360891e1dc363c99c0de5)\n\n**new alerts:**\n\n* 1 for Variable defined multiple times" ]

[]

[]

[]

[]

[]

[ "> \r\n> \r\n> Doesn't this just dump the files into the project dir? @blisc any thoughts if this is correct ?\r\n\r\nActually I made this from \"\" to None last time to hope it fixes the w&b but it made the issue that now we have a folder name \"None\" and then there is another folder under that named \"checkpoints\" where checkpoints are getting stored under. They would not get dumped into the project folder but under a folder named checkpoints.", "Hmm I see, we need it to be None for future code but doing `Path(str(None))' returns \"None\" when we need it to be empty string", "How about explicitly checking for \"None\" and then replacing it with \"\"", "> \r\n> \r\n> How about explicitly checking for \"None\" and then replacing it with \"\"\r\n\r\nDid this this way. Now the version property in the config would be set to None instead of \"\".", "This overall looks like a good solution, @blisc to confirm the approach is correct and merge if its fine." ]

[ { "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" } ]

[]

[ "Thanks for finding this, I'll send a PR tomorrow patching this. If you only need inference, you can freeze the model with model.freeze() and this will prevent any leak. \r\n\r\nIt's the notebook environment which has the potential for leaking cuda tensors, and scripts don't have this leakage.\r\n\r\nThe notebooks are meant to be tutorials, not full scale training scripts. For such scripts please visit the examples/asr/ directory. You might need to modify the scripts to suit finetuning purposes. \r\n\r\nThe model uses batch norm, and is therefore affected by padding inside samples of uneven length. For academic work, we use batch size 1 inference on a single gpu. \r\n\r\nWe do not support TPU training and it is not currently tested. ", "Thanks! I really appreciate the quick response. To clarify, I got the prints before and after to match by adding:\r\n\r\ndel test_batch, log_probs, targets, targets_lengths, encoded_len, greedy_predictions\r\n\r\nBut maybe not all those deletes are needed? I hope that helps!\r\n", "While I have you, I saw in some other comments that \"iter_with_steps\" can help with cuda memory issues. Where would I change that param? Thanks again!", "That argument has been deprecated for quite some time, and it has no impact on memory. If you need to limit training, you can use the PTL trainer argument \"max_steps\"" ]

[]

[]

[ "@okuchaiev fyi regarding the dockerfile ignore and correctly updating the self.trainer inside set_trainer." ]

[]

[]

[]

[]

[]

[]

[ { "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 } ]

[ "Closing. re-open if issue is still relevant." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Thanks Robert for catching the bug! It looks like it has been around for a long time.\r\nWhy do you believe that \"it causes the placement of the masks to be limited to the lowest frequency bins\"?\r\nThe starting point for the frequency masking looks correct to me and can be any frequency. The only problem is that maximum length in frequency dimension was set to 100 instead of 47 (these are the most common values used with this augmentation technique, F=47 and T=100).", "This is an incredible find, thank you @rbracco ! I've verified that your colab (also, very nice way to show the correctness of your fix!) is indeed correct. \r\n\r\nPR approved. ", "Thank you both for the quick review and feedback. \r\n\r\n> Thanks Robert for catching the bug! It looks like it has been around for a long time.\r\n> Why do you believe that \"it causes the placement of the masks to be limited to the lowest frequency bins\"?\r\n> The starting point for the frequency masking looks correct to me and can be any frequency. The only problem is that maximum length in frequency dimension was set to 100 instead of 47 (these are the most common values used with this augmentation technique, F=47 and T=100).\r\n\r\nSorry, that issue only occurs if you have a high F and low T and it's actually the upper frequency bins, but the change fixes the issue anyway.\r\n\r\nAlso, just a heads up `rect_x` here refers to the frequency mask but the frequency is across the y_axis. Since the mask is applied with `x[idx, rect_x : rect_x + w_x, rect_y : rect_y + w_y] = 0.0` and the 2nd dimension is actually the height/feature dimension, and 3rd is width/time so since x refers to the freq dimension in both places and y refers to the time dimension in both, it still works, but if you want to do a quick name change before merging it may make it more clear. Cheers.\r\n\r\n", "A name change would be good, to reflect which axis they effect. Rather than x and y I'd rather just deference them to width and height too if possible. " ]

[]

[]

[ "You generally should not train with such long sequences, since it would use far too much memory + enormous amounts of training data + compute to train a model on long segments. Generally durations of 15-20 second segments are used for training. \r\n\r\nIt would be best to segment the audio in the second dataset. If that is not possible, try to combine both the shorter and longer duration datasets and train for a long time (at least a hundred or more epochs) to fit the long duration audio corpus properly. \r\n\r\nAlso, note that you can use a QuartzNet model trained on 16.7 second audio duration, and evaluate on 2-5 minute long clips with no issue. There may be modest word error rate increase without finetuning it. ", "> You generally should not train with such long sequences, since it would use far too much memory + enormous amounts of training data + compute to train a model on long segments. Generally durations of 15-20 second segments are used for training.\r\n> \r\n> It would be best to segment the audio in the second dataset. If that is not possible, try to combine both the shorter and longer duration datasets and train for a long time (at least a hundred or more epochs) to fit the long duration audio corpus properly.\r\n> \r\n> Also, note that you can use a QuartzNet model trained on 16.7 second audio duration, and evaluate on 2-5 minute long clips with no issue. There may be modest word error rate increase without finetuning it.\r\n\r\nThanks" ]

[]

[]

[ "Thanks for the PR, could you sign your commit according to the instructions on the dco bot ?(https://github.com/NVIDIA/NeMo/pull/2199/checks?check_run_id=2563046850)\r\n", "@titu1994 done with the signoff", "Thanks for the fix !" ]

[]

[]

[]

[]

[]

[ "wouldn't this break all the checkpoints trained with pre_ln=True?", "This pull request **introduces 1 alert** when merging 06d1bb03d4144ea1d95b1656083db015711489c2 into d1b37ed30a70bff00df4778b55b020d8bab2e6fb - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-72c6d188f1fb006ed4703c0de0401a0abd5eb9e6)\n\n**new alerts:**\n\n* 1 for Unused local variable" ]

[]

[]

[]

[]

[]

[ "+ please run tests in an env that does not have pynini installed to ensure pytest tests/ --cpu is runnable " ]

[]

[]

[]

[]

[]

[]

[ { "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": "yzhang123", "id": 4204271, "node_id": "MDQ6VXNlcjQyMDQyNzE=", "avatar_url": "https://avatars.githubusercontent.com/u/4204271?v=4", "gravatar_id": "", "url": "https://api.github.com/users/yzhang123", "html_url": "https://github.com/yzhang123", "followers_url": "https://api.github.com/users/yzhang123/followers", "following_url": "https://api.github.com/users/yzhang123/following{/other_user}", "gists_url": "https://api.github.com/users/yzhang123/gists{/gist_id}", "starred_url": "https://api.github.com/users/yzhang123/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/yzhang123/subscriptions", "organizations_url": "https://api.github.com/users/yzhang123/orgs", "repos_url": "https://api.github.com/users/yzhang123/repos", "events_url": "https://api.github.com/users/yzhang123/events{/privacy}", "received_events_url": "https://api.github.com/users/yzhang123/received_events", "type": "User", "user_view_type": "public", "site_admin": false }, { "login": "borisfom", "id": 14189615, "node_id": "MDQ6VXNlcjE0MTg5NjE1", "avatar_url": "https://avatars.githubusercontent.com/u/14189615?v=4", "gravatar_id": "", "url": "https://api.github.com/users/borisfom", "html_url": "https://github.com/borisfom", "followers_url": "https://api.github.com/users/borisfom/followers", "following_url": "https://api.github.com/users/borisfom/following{/other_user}", "gists_url": "https://api.github.com/users/borisfom/gists{/gist_id}", "starred_url": "https://api.github.com/users/borisfom/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/borisfom/subscriptions", "organizations_url": "https://api.github.com/users/borisfom/orgs", "repos_url": "https://api.github.com/users/borisfom/repos", "events_url": "https://api.github.com/users/borisfom/events{/privacy}", "received_events_url": "https://api.github.com/users/borisfom/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "What are you trying to accomplish? a) just to extract .pt from .nemo or b) export the model to Torchscript ? \r\nIf a): then simply use 'tar xvf model.nemo'\r\nIf b): Megatron model currently is not exportable to Torchscript. You can export it to ONNX though by supplying name with .onnx extension to export()\r\n", "I was trying to do B - export to torchscript. I did end up going the ONNX route." ]

[]

[]

[ "LGTM." ]

[]

[]

[ "This pull request **introduces 1 alert** when merging a5b2e9be03ccadfd608d829ab10809af788c0a2f into b44a1077df4579ecdba9a5386fbfe2c0bcd20667 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-2d1e71128b7ca228053fb661556cd8f6fede1c7f)\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" } ]

[]

[ "I'm not too familiar with the internals of PTL, but I'm guessing the solution is to determine whether there are GPUs _or_ TPUs and not _just_ GPUs.", "We don't support TPU training, and it's not tested. If you get it working, please let us know. ", "looks like you can try disabling learning rate scheduler. but, yes, training on TPU's isn't something we maintain - in G's Colab you should be able to just use GPUs" ]

[]

[]

[]

[]

[ { "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's better to have higher number_of_samples/speaker with shorter durations (of about 4 sec) than vice versa. ", "Hi, @nithinraok Thank you!" ]

[ { "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 should be fixed via https://github.com/NVIDIA/NeMo/pull/2178" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Librispeech KenLM should be built on concatenated datasets of entire Librispeech lm text corpus (after normalization to lower case) + the 3 train transcripts. Overall it should be close to 41 M lines of text rather than the 300k lines of text here. \r\n\r\nSecondly, this sentence is confusing - \"However I got higher WER using LM i.e. approx 16.4% as compared to 3.4% WER using only CitriNet without LM on my audios.\"\r\n\r\nIs the 3.4% with the Citrinet without LM or Citrinet with LM? Did you get 16.4% with an LM on a different dataset or on the same Dataset that you got 3.4%? \r\n\r\nIf it's hour own audio clip, then Citrinet might simply not be good with greedy inference either. Check that greedy wer is close to the 16% you so with LM or higher ", "Thank you for your response.\r\n\r\nI got 3.4% using Citrinet without LM. And got 16.4% using Citrinet with LM.\r\n\r\nI am getting both the above WER on the same data input.\r\n\r\nI used only 1-2 minutes of audios as input.\r\n\r\n\r\n \r\n\r\n\r\n", "The Citrinet Checkpoint available on Nemo is trained on roughly 7000 hours of speech - far more than just Librispeech. So a Librispeec LM (without the actual Librispeech text corpus) will only go so far as to improve or hurt wer. \r\n\r\nCould you try as an experiment what is the result with alpha and betas set to 0?", "Okay, I'll try to train LM on large corpus.\r\n\r\nWhen I set alpha=0 and beta=0, the WER is 3.50%", "Ok so the fact that alpha 0 beta 0 gets close to greedy scores shows that your steps are correct and LM will help - just that it needs proper dataset + hyper parameter search to get good scores. You can tune your LM on the Librispeech dev set. ", "ok, understood. Thank you for your time." ]

[]

[]

[]

[]

[]

[ "This pull request **introduces 1 alert** when merging bc2cdec3c9e9767af3ef755a8160db979f572d59 into d9aab777c4954d13ae95d8b71ef08de65c947f5c - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-30686a7b465ee52e9ed5f4017a818b2b8d5eb153)\n\n**new alerts:**\n\n* 1 for Unused import", "This pull request **introduces 1 alert** when merging 02b4f9f07c51d75a4296fb11fc0cec6df0dd1686 into b7ca1ea0c72752b37351623a1a5ee4d5d1b447e0 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-1f075d3f2d7f3339d882aff37138fa5b7f3fab21)\n\n**new alerts:**\n\n* 1 for Unused import" ]

[]

[]

[]

[]

[]

[ "This pull request **introduces 1 alert** when merging 1c2ead52bc3b9c67f46f46636b31a0fc01a39935 into b44a1077df4579ecdba9a5386fbfe2c0bcd20667 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-49f72b46b780b693b22b9ea54e2714adcf6aa8e4)\n\n**new alerts:**\n\n* 1 for Unused import", "This pull request **introduces 1 alert** when merging af4f0ae39bbcb6b1c3a169c400ddb294bc21e6ad into b44a1077df4579ecdba9a5386fbfe2c0bcd20667 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-fa66ff601ca8f005bd838a12936ec73ea14b545d)\n\n**new alerts:**\n\n* 1 for Unused import", "This pull request **introduces 1 alert** when merging ac3ddf6a8c5b22e732056754a4f3d6a5c1681c2f into b44a1077df4579ecdba9a5386fbfe2c0bcd20667 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-04bad2d21a3d210cc6265712240fb61299f28970)\n\n**new alerts:**\n\n* 1 for Unused import", "The naming convention for the latest checkpoint in lightning changed in 1.3, so I updated it.", "This pull request **introduces 1 alert** when merging b4b5854e2db3a1fe15e57405ca8b95ae2640c937 into b44a1077df4579ecdba9a5386fbfe2c0bcd20667 - [view on LGTM.com](https://lgtm.com/projects/g/NVIDIA/NeMo/rev/pr-700a8bce21ca405f90a531a19b4793ce0b311f18)\n\n**new alerts:**\n\n* 1 for Unused import", "Thanks @blisc !" ]

[]

[]

[ "Thanks for noticing this bug. I'll fix this in #2166 ", "Oh, it looks like someone already fixed it. We will fix this in the upcoming 1.0 release." ]

[]

[]

[ "If it's the KenLM binary causing the memory issue, then a smaller N-gram can be used. However we don't have much control over the c++ swig decoders so from our side there isn't a sufficient solution to reduce memory. Maybe @vsl9 has some ideas?", "If the memory issue is from the beam search decoder, then you may try to call the beam search decoder on each sample separately. So divide your batch into smaller batches or even single samples before calling the decoder. Beam search decoding is mostly done on CPU and batch computation does not increase the speed of the decoder, so you can get the same speed with single samples. \r\nOther thing you can try is to lower the number for num_cpus, not sure how much it helps.", "> If the memory issue is from the beam search decoder, then you may try to call the beam search decoder on each sample separately. So divide your batch into smaller batches or even single samples before calling the decoder. Beam search decoding is mostly done on CPU and batch computation does not increase the speed of the decoder, so you can get the same speed with single samples.\r\n> Other thing you can try is to lower the number for num_cpus, not sure how much it helps.\r\n\r\nThanks for the suggestions. I am trying to use it on on single instance so my batch size is 1. Ram usage increases about 2.5 GBs when I am instantiating BeamSearchDecoderWithLM. So this is before even feeding anything to the network.", "I see. In that case, memory consumption is mainly due to the ngram model. You may follow Som's suggestion as ngram models need significant memory. That is one of the main reasons they are not preferable for devices with limited resources. Using a model with smaller vocabulary size can also help to reduce the size of the ngram model but not sure if it solves your issue completely.", "Is this the Cuda storage? Or just normal RAM? Thanks!", "> Is this the Cuda storage? Or just normal RAM? Thanks!\r\n\r\nI am only using normal RAM. My system doesn't have GPU.", "Yes, it is CPU RAM usage for n-gram LM. Scorer object is instantiated during BeamSearchDecoderWithLM construction. And it loads n-gram LM into RAM. Reducing size of LM (number of tokens/words, LM's order) should help." ]

[]

[]

[ "> exortable changes look good to me. @blisc what about tts changes?\r\n\r\ntts changes are to conform to new extended syntax accepted by onnx.export (optional Dict as the last input tuple element, representing kwargs) instead of former non-standard extended syntax I briefly introduced in previous revision (accepting a Dict instead of a tuple as inputs, not really used except in this test example). " ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "@okuchaiev @ericharper we need to enable the full CI before merging this", "> @okuchaiev @ericharper we need to enable the full CI before merging this\r\n\r\nLooks like the Jenkinsfile reverted to main. I replaced with v1.0.0 again." ]

[]

[]

[ "@titu1994 Could you review the register_artifact docs?\r\n\r\n@MaximumEntropy Could you review NLP/NMT docs?\r\n\r\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]