Per the competition rules, post links to your external data sources here before the deadline specified. Once it has been posted, you do not need to post it again.
", "messages": [ { "id": 887433, "postDate": "2020-06-15T17:08:18.173Z", "content": "Per the competition rules, post links to your external data sources here before the deadline specified. Once it has been posted, you do not need to post it again.
", "rawMarkdown": "Per the competition rules, post links to your external data sources here before the deadline specified. Once it has been posted, you do not need to post it again.", "votes": 24 }, { "id": 887538, "postDate": "2020-06-15T18:17:35.543Z", "content": "Do we need additional permissions or licenses from individual birds appearing in any external dataset? (DFDC reference).
", "rawMarkdown": "Do we need additional permissions or licenses from individual birds appearing in any external dataset? (DFDC reference). ", "votes": 18 }, { "id": 911336, "postDate": "2020-07-01T17:02:35.620Z", "content": "In addition to my post earlier, here's the download link for the 2020 BirdCLEF validation data (which contains 60 minutes of soundscapes from North America and Peru): https://drive.google.com/file/d/1OIEJ9W6W5MNMXOicjiKivGmxTtskcJOn/view?usp=sharing
\n\nYou can use this data to test the performance of your system (of course, you'd have to adapt the ground truth to fit your needs). It is not permitted to use any other BirdCLEF data from current or past editions.
\n\nLet me know if you have any issues with the dataset.
", "rawMarkdown": "In addition to my post earlier, here's the download link for the 2020 BirdCLEF validation data (which contains 60 minutes of soundscapes from North America and Peru): https://drive.google.com/file/d/1OIEJ9W6W5MNMXOicjiKivGmxTtskcJOn/view?usp=sharing\n\nYou can use this data to test the performance of your system (of course, you'd have to adapt the ground truth to fit your needs). It is not permitted to use any other BirdCLEF data from current or past editions. \n\nLet me know if you have any issues with the dataset.", "votes": 14, "replies": [ { "id": 911528, "postDate": "2020-07-01T19:57:44.560Z", "content": "Is the training data of birdclef 2020 ok to use ? it is 70k clips from xeno canto.
", "rawMarkdown": "Is the training data of birdclef 2020 ok to use ? it is 70k clips from xeno canto." }, { "id": 912040, "postDate": "2020-07-02T07:09:33.317Z", "content": "Well, not sure if it is going to be much of a help - it contains a lot of species that are irrelevant for this competition and it should contain the same recordings for species that are relevant for this competition. So I guess, yes, you can take a look but make sure to post it here if you find any portion of the data helpful.
", "rawMarkdown": "Well, not sure if it is going to be much of a help - it contains a lot of species that are irrelevant for this competition and it should contain the same recordings for species that are relevant for this competition. So I guess, yes, you can take a look but make sure to post it here if you find any portion of the data helpful." }, { "id": 912709, "postDate": "2020-07-02T17:04:59.243Z", "content": ">it contains a lot of species that are irrelevant for this competition
\n\nYes... I confirmed type of species in BirdCLEF2020.\nPER recordings do not include any spices of this competition. Meanwhile, SSW recordings have 76 types of birds, but only 17 types of them match to this competition.\nI think if test datasets include all 264 species, BirdCLEF2020 dataset may not be sufficient to evaluate our models...
\n\n@stefankahl \nIf I'm wrong, please correct me.
\n\ncomments added,
\n\nAnd ground truth file is messy.\nSome rows are duplicated,\ne.g.\nSSW52_20170429 file has 3 duplicated rows.\nOne of them is like this,\n00:02:30-00:02:35 purfin\n00:02:30-00:02:35 purfin
", "rawMarkdown": ">it contains a lot of species that are irrelevant for this competition\n\nYes... I confirmed type of species in BirdCLEF2020.\nPER recordings do not include any spices of this competition. Meanwhile, SSW recordings have 76 types of birds, but only 17 types of them match to this competition.\nI think if test datasets include all 264 species, BirdCLEF2020 dataset may not be sufficient to evaluate our models...\n\n@stefankahl \nIf I'm wrong, please correct me.\n\n\n---\n\ncomments added,\n\nAnd ground truth file is messy.\nSome rows are duplicated,\ne.g.\nSSW52_20170429 file has 3 duplicated rows.\nOne of them is like this,\n00:02:30-00:02:35\tpurfin\n00:02:30-00:02:35\tpurfin", "votes": 2 }, { "id": 913586, "postDate": "2020-07-03T09:44:49.703Z", "content": "I manually looked at the BirdCLEF validation data and it seems that it contains labels for 16 species (I may have missed one) and all of them seem to be part of the training data for this competition. Not sure why the GT contains duplicates but it is save to ignore them. Sure, PER soundscapes are of no use but the SSW files should be a nice way to validate (yet, they are not the busiest soundscapes but suppressing false positives is key!).
\n\nAnd just as a reminder: We chose the training data according to lists of birds that MIGHT occur at the recording sites of the test data - not knowing which birds will actually vocalize is part of the challenge we're facing when deploying recorders.
", "rawMarkdown": "I manually looked at the BirdCLEF validation data and it seems that it contains labels for 16 species (I may have missed one) and all of them seem to be part of the training data for this competition. Not sure why the GT contains duplicates but it is save to ignore them. Sure, PER soundscapes are of no use but the SSW files should be a nice way to validate (yet, they are not the busiest soundscapes but suppressing false positives is key!). \n\nAnd just as a reminder: We chose the training data according to lists of birds that MIGHT occur at the recording sites of the test data - not knowing which birds will actually vocalize is part of the challenge we're facing when deploying recorders.", "votes": 6 }, { "id": 914431, "postDate": "2020-07-03T21:32:26.383Z", "content": "There are 76 labels in the validation set of Birdclef 2020, 17 of which are in this competition data.
", "rawMarkdown": "There are 76 labels in the validation set of Birdclef 2020, 17 of which are in this competition data." }, { "id": 914782, "postDate": "2020-07-04T08:17:20.203Z", "content": "Well, I guess if you combine PER and SSW that's the case. But again, only the SSW soundscapes are of relevance for this competition, PER recordings are from South America - hence the lack of label overlap. All species annotated in the SSW files should be part of the training data for this competition and these soundscapes can be used for validation. We can't provide you with any other soundscape data since annotations are so hard to come by. Yet, optimizing the false positive rate of your classifier should be possible with the SSW soundscapes.
", "rawMarkdown": "Well, I guess if you combine PER and SSW that's the case. But again, only the SSW soundscapes are of relevance for this competition, PER recordings are from South America - hence the lack of label overlap. All species annotated in the SSW files should be part of the training data for this competition and these soundscapes can be used for validation. We can't provide you with any other soundscape data since annotations are so hard to come by. Yet, optimizing the false positive rate of your classifier should be possible with the SSW soundscapes.", "votes": 3 }, { "id": 914995, "postDate": "2020-07-04T12:08:43.280Z", "content": "Hello,
\n\nThanks for clarifying this, i missed the location information.
", "rawMarkdown": "Hello,\n\nThanks for clarifying this, i missed the location information.\n" }, { "id": 925174, "postDate": "2020-07-11T21:31:14.430Z", "content": "@stefankahl Thanks for the link. Can we use this data for the training part ?
", "rawMarkdown": "@stefankahl Thanks for the link. Can we use this data for the training part ? " }, { "id": 959531, "postDate": "2020-08-05T17:04:14.193Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 893651, "postDate": "2020-06-19T19:10:20.760Z", "content": "Additional recordings from https://www.xeno-canto.org/
\nSome are discussed here and maintained as Kaggle datasets with appropriate per-record license:
https://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition
\nhttps://github.com/qiuqiangkong/audioset_tagging_cnn
ResNeSt pre-trained weights using in my baseline
\nhttps://github.com/zhanghang1989/ResNeSt
pytorch-pfn-extras (for training)
\nhttps://github.com/pfnet/pytorch-pfn-extras
previous kaggle challenges\n* https://www.kaggle.com/c/mlsp-2013-birds/data\n* https://www.kaggle.com/c/multilabel-bird-species-classification-nips2013/data
", "rawMarkdown": "previous kaggle challenges\n* https://www.kaggle.com/c/mlsp-2013-birds/data\n* https://www.kaggle.com/c/multilabel-bird-species-classification-nips2013/data", "votes": 6, "replies": [ { "id": 903610, "postDate": "2020-06-27T01:45:32.180Z", "content": "niiiiiiiiice
", "rawMarkdown": "niiiiiiiiice" } ] }, { "id": 914485, "postDate": "2020-07-04T00:47:06.470Z", "content": "Environmental sound and noise sound for nocall.\nhttps://www.youtube.com/watch?v=xNN7iTA57jM\nhttps://www.youtube.com/watch?v=8plwv25NYRo\nhttps://www.youtube.com/watch?v=lR4GNWcwAI8\nhttps://www.youtube.com/watch?v=4KzFe50RQkQ
", "rawMarkdown": "Environmental sound and noise sound for nocall.\nhttps://www.youtube.com/watch?v=xNN7iTA57jM\nhttps://www.youtube.com/watch?v=8plwv25NYRo\nhttps://www.youtube.com/watch?v=lR4GNWcwAI8\nhttps://www.youtube.com/watch?v=4KzFe50RQkQ", "votes": 3, "replies": [ { "id": 996135, "postDate": "2020-09-03T05:37:23.767Z", "content": "haha dude that audio is FULL of birdcalls!
", "rawMarkdown": "haha dude that audio is FULL of birdcalls!" } ] }, { "id": 953117, "postDate": "2020-07-31T14:44:48.177Z", "content": "[Animal Sound Archive] ()https://www.gbif.org/dataset/b7ec1bf8-819b-11e2-bad2-00145eb45e9a) Published by Museum für Naturkunde Berlin
\n\nThe Animal Sound Archive at the Museum fuer Naturkunde Berlin (German: Tierstimmenarchiv) is one of the oldest and largest worldwide. Founded in 1951 by Professor Guenter Tembrock the collection consists now of around 130 000 records of animal voices.
", "rawMarkdown": "[Animal Sound Archive] ()https://www.gbif.org/dataset/b7ec1bf8-819b-11e2-bad2-00145eb45e9a) Published by Museum für Naturkunde Berlin\n\nThe Animal Sound Archive at the Museum fuer Naturkunde Berlin (German: Tierstimmenarchiv) is one of the oldest and largest worldwide. Founded in 1951 by Professor Guenter Tembrock the collection consists now of around 130 000 records of animal voices.", "votes": 4 }, { "id": 943332, "postDate": "2020-07-24T09:52:22.280Z", "content": "Microphone wind noise simulator\nhttps://github.com/kenders2000/MicWindNoiseGenerator
", "rawMarkdown": "Microphone wind noise simulator\nhttps://github.com/kenders2000/MicWindNoiseGenerator", "votes": 4 }, { "id": 935200, "postDate": "2020-07-19T06:30:17.233Z", "content": "I just learned about this very interesting resource - radio aporee. It is an enormous collection of soundscape recordings from all over the world! 🙂 Might be very useful to see how our models perform on data coming from various devices.
", "rawMarkdown": "I just learned about this very interesting resource - [radio aporee](https://aporee.org/maps/info/#archive). It is an enormous collection of soundscape recordings from all over the world! 🙂 Might be very useful to see how our models perform on data coming from various devices.", "votes": 4 }, { "id": 898238, "postDate": "2020-06-23T11:49:58.650Z", "content": "I was looking for soundscape recordings and found the LifeCLEF Bird challenges.
\n\n\n\n\nThe test data consists of 153 soundscapes recorded in Peru, the USA, and Germany. Each soundscape is of ten-minute duration and contains high quantities of (overlapping) bird vocalizations.
\n
I am bit worried as the two test sets might overlap.
\n\n\n", "rawMarkdown": "I was looking for soundscape recordings and found the LifeCLEF Bird challenges. \n\n* https://www.aicrowd.com/challenges/lifeclef-2020-bird-monophone\n* https://www.aicrowd.com/challenges/lifeclef-2018-bird-monophone\n\n> The test data consists of 153 soundscapes recorded in Peru, the USA, and Germany. Each soundscape is of ten-minute duration and contains high quantities of (overlapping) bird vocalizations.\n\nI am bit worried as the two test sets might overlap.\n\n> The hidden test_audio directory contains approximately 150 recordings in mp3 format, each roughly 10 minutes long.\n\n\n\n\n\n\n", "votes": 4, "replies": [ { "id": 898243, "postDate": "2020-06-23T11:53:32Z", "content": "The hidden test_audio directory contains approximately 150 recordings in mp3 format, each roughly 10 minutes long.
\n
I see you are trying hard to find the test data 😁 \nBut this test data looks eerily similar.
", "rawMarkdown": "I see you are trying hard to find the test data 😁 \nBut this test data looks eerily similar.", "votes": 1 }, { "id": 898253, "postDate": "2020-06-23T11:58:01.557Z", "content": "I found the CLEF competitions by clicking twice starting from this discussion thread :)
\n\nSince I found them it is better to share early.\nAnyway I would like to create useful local validation set with soundscape recordings. I don't like when local CV does not match LB...
", "rawMarkdown": "I found the CLEF competitions by clicking twice starting from this discussion thread :)\n\nSince I found them it is better to share early.\nAnyway I would like to create useful local validation set with soundscape recordings. I don't like when local CV does not match LB...", "votes": 4 }, { "id": 898655, "postDate": "2020-06-23T16:54:12.127Z", "content": "The LifeClef files appear similar because some of the same organizations were involved in generating the datasets used by each competition. However, there is no overlap between the files.
", "rawMarkdown": "The LifeClef files appear similar because some of the same organizations were involved in generating the datasets used by each competition. However, there is no overlap between the files.", "votes": 6 }, { "id": 898667, "postDate": "2020-06-23T17:01:38.913Z", "content": "So we are allowed to use them (even manually annotate & train models on them), right?
", "rawMarkdown": "So we are allowed to use them (even manually annotate & train models on them), right?", "votes": 1 }, { "id": 899494, "postDate": "2020-06-24T09:18:10.263Z", "content": "Here's my concern: The LifeCLEF dataset is not explicitly public - you have to register to get the files and they will be gone once the second round of submissions is over. We have to ensure that this actually qualifies as a valid external source according to the rules. We will look into that and keep you posted.
", "rawMarkdown": "Here's my concern: The LifeCLEF dataset is not explicitly public - you have to register to get the files and they will be gone once the second round of submissions is over. We have to ensure that this actually qualifies as a valid external source according to the rules. We will look into that and keep you posted." }, { "id": 899523, "postDate": "2020-06-24T09:36:11.487Z", "content": "@stefankahl thanks for looking into that, I saw you are also an organizer in those competitions. Btw I could register and download the closed (2018) competition data as well. If we are allowed I am happy to publish the datasets here at kaggle although could not find any licensing terms in the rules.
", "rawMarkdown": "@stefankahl thanks for looking into that, I saw you are also an organizer in those competitions. Btw I could register and download the closed (2018) competition data as well. If we are allowed I am happy to publish the datasets here at kaggle although could not find any licensing terms in the rules. \n", "votes": 1 }, { "id": 899696, "postDate": "2020-06-24T11:53:31.823Z", "content": "The 2018 data does not contain any North American soundscapes (and no labels). The 2020 dataset would be more interesting, although I think we should avoid to allow training on soundscape data (even though there's no overlap with the Kaggle test set, it just wouldn't fit the task). But it would be a nice validation dataset - we will discuss that and let you know.
", "rawMarkdown": "The 2018 data does not contain any North American soundscapes (and no labels). The 2020 dataset would be more interesting, although I think we should avoid to allow training on soundscape data (even though there's no overlap with the Kaggle test set, it just wouldn't fit the task). But it would be a nice validation dataset - we will discuss that and let you know.", "votes": 5 }, { "id": 902798, "postDate": "2020-06-26T11:16:21.950Z", "content": "Having a validation set could be really useful for figuring out how to go from training a classifier on train data to predicting on soundscape. Right now we are essentially flying blind in this regard.
\n\nI am also curious, @stefankahl would you by any chance be aware of a resource that would list birds along with how common they are to certain regions, whether they are transitory, etc? Would be ideal if this contained an ebird code but if not that is not a problem. I looked on ebird.org but maybe I was looking in the wrong place or this information does not exist there.
\n\nThanks a lot for all your help!
", "rawMarkdown": "Having a validation set could be really useful for figuring out how to go from training a classifier on train data to predicting on soundscape. Right now we are essentially flying blind in this regard.\n\nI am also curious, @stefankahl would you by any chance be aware of a resource that would list birds along with how common they are to certain regions, whether they are transitory, etc? Would be ideal if this contained an ebird code but if not that is not a problem. I looked on ebird.org but maybe I was looking in the wrong place or this information does not exist there.\n\nThanks a lot for all your help!", "votes": 3 }, { "id": 903603, "postDate": "2020-06-27T01:27:55.200Z", "content": "@stefankahl I think some kind of soundscape data including multiple birdcalls would confer a huge benefit to this competition and improve the resulting models since it would allow participants to estimate the strength of their models (and certain important hyperparameters like prediction score threshold) WITHOUT mercilessly probing the public test set.
\n\nJust my two cents.
", "rawMarkdown": "@stefankahl I think some kind of soundscape data including multiple birdcalls would confer a huge benefit to this competition and improve the resulting models since it would allow participants to estimate the strength of their models (and certain important hyperparameters like prediction score threshold) WITHOUT mercilessly probing the public test set. \n\nJust my two cents.", "votes": 4 }, { "id": 903607, "postDate": "2020-06-27T01:37:13.487Z", "content": "\n\n\nThe 2018 data does not contain any North American soundscapes
\n
ha-HA! so all 3 sites are in North America! We're on to you Stefan.
", "rawMarkdown": "> The 2018 data does not contain any **North American** soundscapes\n\nha-HA! so all 3 sites are in North America! We're on to you Stefan." }, { "id": 911073, "postDate": "2020-07-01T14:53:09.127Z", "content": "Hm, not sure if this was a secret. The training data only contains North American species :)
", "rawMarkdown": "Hm, not sure if this was a secret. The training data only contains North American species :)", "votes": 2 }, { "id": 911089, "postDate": "2020-07-01T15:03:30.330Z", "content": "Maybe it is worth mentioning on the overview/data page. I only found this info in the forum.
", "rawMarkdown": "Maybe it is worth mentioning on the overview/data page. I only found this info in the forum.", "votes": 1 }, { "id": 911091, "postDate": "2020-07-01T15:04:02.683Z", "content": "I talked to the other hosts about the use of BirdCLEF data for validation and we decided that this is not going to be problematic. Therefore, I put the 2020 validation data on Google Drive so you don't have to register for BirdCLEF (unless, of course, you like to submit, the challenge is still open for post-deadline submissions). In any case, the dataset contains 60 minutes of soundscapes from Peru (probably not of much use for this competition) and 60 minutes from North America (no overlap with the Kaggle test data, but as you mentioned it could be valuable to test your systems). I think the Kaggle training data should cover all species from the BirdCLEF validation data, but the contained soundscapes aren't the busiest ones. I know I can't prevent you from training on soundscape data, but I would like to remind you of the goal of this competition: Train on focal recordings and apply to soundscapes - that's the only way that we can apply a detection system to a new location without annotating hours of soundscapes first. Due to this, the use of other data from BirdCLEF will not be permitted (please do not annotate other soundscapes, that would be against the competition's intent).
\n\nHere's the download link: https://drive.google.com/file/d/1OIEJ9W6W5MNMXOicjiKivGmxTtskcJOn/view?usp=sharing
\n\nLet me know if you have any issues with the dataset.
", "rawMarkdown": "I talked to the other hosts about the use of BirdCLEF data for validation and we decided that this is not going to be problematic. Therefore, I put the 2020 validation data on Google Drive so you don't have to register for BirdCLEF (unless, of course, you like to submit, the challenge is still open for post-deadline submissions). In any case, the dataset contains 60 minutes of soundscapes from Peru (probably not of much use for this competition) and 60 minutes from North America (no overlap with the Kaggle test data, but as you mentioned it could be valuable to test your systems). I think the Kaggle training data should cover all species from the BirdCLEF validation data, but the contained soundscapes aren't the busiest ones. I know I can't prevent you from training on soundscape data, but I would like to remind you of the goal of this competition: Train on focal recordings and apply to soundscapes - that's the only way that we can apply a detection system to a new location without annotating hours of soundscapes first. Due to this, the use of other data from BirdCLEF will not be permitted (please do not annotate other soundscapes, that would be against the competition's intent).\n\nHere's the download link: https://drive.google.com/file/d/1OIEJ9W6W5MNMXOicjiKivGmxTtskcJOn/view?usp=sharing\n\nLet me know if you have any issues with the dataset.", "votes": 8 }, { "id": 911120, "postDate": "2020-07-01T15:21:22.170Z", "content": "Regarding good resources on bird species: I think eBird.org is a good starting point. They have range maps for all species and so-called \"bar charts\" where you can see the levels of abundance for a given location. (e.g., for the American Robin you can browse to https://ebird.org/species/amerob). But I also think that the recording locations and dates in the training data should reflect species abundance across all seasons and you should be able to figure out if a species is migrating or not without manual interference. The number of recordings (in the extended training data) and diversity of recording locations should also reflect how common a species is.
\n\nBTW: eBird also has a nice public API (https://documenter.getpostman.com/view/664302/S1ENwy59?version=latest), yet I am not sure if any of the features it offers would be suited to help you for this competition.
\n\nAnd again, as a reminder: Please do not crawl these sites without considering the load you might impose on the servers.
", "rawMarkdown": "Regarding good resources on bird species: I think eBird.org is a good starting point. They have range maps for all species and so-called \"bar charts\" where you can see the levels of abundance for a given location. (e.g., for the American Robin you can browse to https://ebird.org/species/amerob). But I also think that the recording locations and dates in the training data should reflect species abundance across all seasons and you should be able to figure out if a species is migrating or not without manual interference. The number of recordings (in the extended training data) and diversity of recording locations should also reflect how common a species is. \n\nBTW: eBird also has a nice public API (https://documenter.getpostman.com/view/664302/S1ENwy59?version=latest), yet I am not sure if any of the features it offers would be suited to help you for this competition.\n\nAnd again, as a reminder: Please do not crawl these sites without considering the load you might impose on the servers.", "votes": 3 }, { "id": 911171, "postDate": "2020-07-01T15:44:33.277Z", "content": "@stefankahl Thanks a lot for the updates and for sharing the data. Though I strongly recommended to post these comments / data as a separate pinned topic (like the XC-crawling one) or part of the official rules / official data, so that it is clearly visible to all competitors (rather than lost in the middle of this thread) especially considering some of these points affect the rules of competing / modelling in this competition.\ncc: @sohier
", "rawMarkdown": "@stefankahl Thanks a lot for the updates and for sharing the data. Though I strongly recommended to post these comments / data as a separate pinned topic (like the XC-crawling one) or part of the official rules / official data, so that it is clearly visible to all competitors (rather than lost in the middle of this thread) especially considering some of these points affect the rules of competing / modelling in this competition.\ncc: @sohier ", "votes": 1 }, { "id": 911342, "postDate": "2020-07-01T17:06:33.617Z", "content": "I added another post that contains the link. However, I think this is the right place to announce it since it is additional data which is not part of the official dataset (despite being announced by a host). But of course, feel free to post another thread or notebook that explores the recordings and maybe even manages to provide a ground truth that fits this competition (right now, it's BirdCLEF-specific).
", "rawMarkdown": "I added another post that contains the link. However, I think this is the right place to announce it since it is additional data which is not part of the official dataset (despite being announced by a host). But of course, feel free to post another thread or notebook that explores the recordings and maybe even manages to provide a ground truth that fits this competition (right now, it's BirdCLEF-specific).", "votes": 4 }, { "id": 911641, "postDate": "2020-07-01T22:37:10.527Z", "rawMarkdown": "", "isDeleted": true }, { "id": 916583, "postDate": "2020-07-05T19:46:33.150Z", "rawMarkdown": "", "isDeleted": true }, { "id": 963132, "postDate": "2020-08-08T17:20:19.673Z", "content": "@stefankahl You said test set contains only North America species correct? If I'm not wrong there are 262 species over 264 covering North America in train set. There are 2 species never seen in North America:\n- gnwtea: Eurasian Teal\n- hergul: European Herring Gull
\n\nIs that correct? If so, we can label them as \"nocall\" on inference time if there are detected.\nIt won't make a big difference but it's good to know.
", "rawMarkdown": "@stefankahl You said test set contains only **North America** species correct? If I'm not wrong there are 262 species over 264 covering [North America](https://www.countries-ofthe-world.com/countries-of-north-america.html) in train set. There are 2 species never seen in North America:\n- gnwtea: [Eurasian Teal](https://www.xeno-canto.org/species/Anas-crecca)\n- hergul: [European Herring Gull](https://www.xeno-canto.org/species/Larus-argentatus)\n\nIs that correct? If so, we can label them as \"**nocall**\" on inference time if there are detected.\nIt won't make a big difference but it's good to know.", "votes": 1 }, { "id": 965399, "postDate": "2020-08-10T15:46:51.127Z", "content": "The test set recording locations are in America. Birds occasionally take a wrong turn during a migration or get blown about in a storm and end up on the wrong continent so I would guess that the recordings of those two species in the train set were of rare sightings in North America. Birds seen outside of their usual territories are very popular with birders, it's not too surprising that that they would be recorded.
", "rawMarkdown": "The test set recording locations are in America. Birds occasionally take a wrong turn during a migration or get blown about in a storm and end up on the wrong continent so I would guess that the recordings of those two species in the train set were of rare sightings in North America. Birds seen outside of their usual territories are very popular with birders, it's not too surprising that that they would be recorded.", "votes": 3 }, { "id": 965493, "postDate": "2020-08-10T17:10:25.847Z", "content": "Couple things could be going on here... a) We're using the eBird codes, rather than the 6-letter codes. They MOSTLY agree, except for where they don't. b) Both of these birds have American and Eurasian flavors.
\n\ngnwtea == Green Winged Teal\nhergul == Herring Gull
", "rawMarkdown": "Couple things could be going on here... a) We're using the eBird codes, rather than the 6-letter codes. They MOSTLY agree, except for where they don't. b) Both of these birds have American and Eurasian flavors.\n\ngnwtea == [Green Winged Teal](https://www.allaboutbirds.org/guide/Green-winged_Teal/maps-range)\nhergul == [Herring Gull](https://www.allaboutbirds.org/guide/Herring_Gull/maps-range)", "votes": 2 }, { "id": 966159, "postDate": "2020-08-11T07:41:44.830Z", "content": "wrt gnwtea and hergul, I did some checking on the location info in train and none were from North America. however, figured it was best to retain them just in case of migration timing not represented in train or other bird misadventures.
\n\nfor external data would like to add \nhttps://github.com/rwightman/pytorch-image-models
", "rawMarkdown": "wrt gnwtea and hergul, I did some checking on the location info in train and none were from North America. however, figured it was best to retain them just in case of migration timing not represented in train or other bird misadventures. \n\nfor external data would like to add \nhttps://github.com/rwightman/pytorch-image-models" } ] }, { "id": 1006880, "postDate": "2020-09-11T15:58:05.493Z", "content": "https://www.kaggle.com/ludovick/xenoexternalwav0
\nhttps://www.kaggle.com/ludovick/xenoexternalwav1
EfficientNet-PyTorch:https://github.com/lukemelas/EfficientNet-PyTorch
\nPANN https://github.com/qiuqiangkong/audioset_tagging_cnn/
\nGoogle AudioSet https://research.google.com/audioset/download.html
\nfastai and pre-trained models: https://github.com/fastai/fastai
\nPretrained models for Pytorch https://github.com/cadene/pretrained-models.pytorch
\ntorchvision models and pretrained weights https://pytorch.org/docs/stable/torchvision/models.html
\nESC-50 https://github.com/karolpiczak/ESC-50
\nkeras https://keras.io/
\nkeras pre-trained models https://keras.io/api/applications/
Some human speech and general sounds for NN pretraining:
\nhttps://commonvoice.mozilla.org/en/datasets
\nhttps://www.openslr.org/17/
\nhttps://www.openslr.org/28/
Are optimizers and schedulers part of the External Data Thread and need to be reported? Have not seen that kind of post before in any other competition, like for AdamW,Radam,schedulers,dataloaders that are not built-in, so it's good to know.
", "rawMarkdown": "Are optimizers and schedulers part of the External Data Thread and need to be reported? Have not seen that kind of post before in any other competition, like for AdamW,Radam,schedulers,dataloaders that are not built-in, so it's good to know.", "votes": 1, "replies": [ { "id": 990406, "postDate": "2020-08-29T15:26:06.523Z", "content": "I am not sure if organizers though of all these details… It would be great if kaggle established a few version of rules that competons had to choose one. Think competitionright like copyright and authorship right.
", "rawMarkdown": "I am not sure if organizers though of all these details... It would be great if kaggle established a few version of rules that competons had to choose one. Think competitionright like copyright and authorship right." }, { "id": 995860, "postDate": "2020-09-02T21:52:55.320Z", "content": "This is my second competition with exactly the same question in the same topic, did not get an answer then but did not repeat the question, but now I start to wonder, should I tag one from the team to get an answer or how does it work.
", "rawMarkdown": "This is my second competition with exactly the same question in the same topic, did not get an answer then but did not repeat the question, but now I start to wonder, should I tag one from the team to get an answer or how does it work. " }, { "id": 998011, "postDate": "2020-09-04T12:21:00.967Z", "content": "@addisonhoward grateful for an answer when time allows.
", "rawMarkdown": "@addisonhoward grateful for an answer when time allows." }, { "id": 1003058, "postDate": "2020-09-08T16:23:03.687Z", "content": "The competition hosts are ultimately the ones who are responsible for the competition rules, however the external data rule is designed, in spirit, to prevent someone with a \"silver bullet\" dataset to end up winning simply because they had access that nobody else did. Optimizers and schedulers aren't a separate dataset, but are a part of your workflow, and do not require disclosure.
", "rawMarkdown": "The competition hosts are ultimately the ones who are responsible for the competition rules, however the external data rule is designed, in spirit, to prevent someone with a \"silver bullet\" dataset to end up winning simply because they had access that nobody else did. Optimizers and schedulers aren't a separate dataset, but are a part of your workflow, and do not require disclosure.", "votes": 1 } ] }, { "id": 979621, "postDate": "2020-08-21T02:04:12.503Z", "content": "jhartquist's fastai_audio:
\nhttps://github.com/jhartquist/fastai_audio
mogwai's fastai_audio:
\nhttps://github.com/mogwai/fastai_audio
fire:
\nhttps://pypi.org/project/fire/
fastai2_audio:
\nhttps://github.com/rbracco/fastai2_audio
EfficientNet-PyTorch:
\nhttps://github.com/lukemelas/EfficientNet-PyTorch
fastai:
\nhttps://github.com/fastai/fastai
fastbook:
\nhttps://pypi.org/project/fastbook/
resnet18:
\nhttps://download.pytorch.org/models/resnet18-5c106cde.pth
kapre:
\nhttps://github.com/keunwoochoi/kapre
XC463492 · 大山雀 · Parus major:
\nhttps://www.xeno-canto.org/463492
XC464650 · 大山雀 · Parus major:
\nhttps://www.xeno-canto.org/464650
Can I use data from this challenge? https://www.kaggle.com/c/freesound-audio-tagging-2019/data
", "rawMarkdown": "Can I use data from this challenge? https://www.kaggle.com/c/freesound-audio-tagging-2019/data", "votes": 1 }, { "id": 968344, "postDate": "2020-08-12T22:57:03.753Z", "content": "https://www.kaggle.com/anjum48/pytorch-lightning
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://github.com/qiuqiangkong/audioset_tagging_cnn
\nhttps://pytorch.org/docs/stable/torchvision/models.html
Environmental Sound Classification 50
", "rawMarkdown": "[Environmental Sound Classification 50](https://www.kaggle.com/mmoreaux/environmental-sound-classification-50)", "votes": 1 }, { "id": 946130, "postDate": "2020-07-26T11:27:00.380Z", "content": "Pls confirm if BirdVox data can be used for this competition
\n\nhttps://wp.nyu.edu/birdvox/birdvox-full-night/
", "rawMarkdown": "Pls confirm if BirdVox data can be used for this competition\n\nhttps://wp.nyu.edu/birdvox/birdvox-full-night/", "votes": 1, "replies": [ { "id": 946165, "postDate": "2020-07-26T12:04:33.597Z", "content": "BirdVox data can be used, but be aware that night flight calls are often fundamentally different from the types of vocalizations recorded in the train and test data.
", "rawMarkdown": "BirdVox data can be used, but be aware that night flight calls are often fundamentally different from the types of vocalizations recorded in the train and test data.", "votes": 5 } ] }, { "id": 940271, "postDate": "2020-07-22T19:40:50.247Z", "content": "https://github.com/AgaMiko/bird-recognition-review\nhttp://dcase.community/challenge2018/task-bird-audio-detection
", "rawMarkdown": "https://github.com/AgaMiko/bird-recognition-review\nhttp://dcase.community/challenge2018/task-bird-audio-detection", "votes": 1 }, { "id": 918196, "postDate": "2020-07-07T04:26:19.620Z", "content": "Pretrained models for Pytorch\nhttps://github.com/cadene/pretrained-models.pytorch
", "rawMarkdown": "Pretrained models for Pytorch\n[https://github.com/cadene/pretrained-models.pytorch](https://github.com/cadene/pretrained-models.pytorch)", "votes": 1 }, { "id": 909771, "postDate": "2020-06-30T20:13:06.150Z", "content": "", "rawMarkdown": "https://ebird.org/", "votes": 1 }, { "id": 895902, "postDate": "2020-06-21T17:16:19.917Z", "content": "All 3 training and 3 validation datasets for this challenge.\nhttp://dcase.community/challenge2018/task-bird-audio-detection#audio-datasets
\n\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019
", "rawMarkdown": "All 3 training and 3 validation datasets for this challenge.\nhttp://dcase.community/challenge2018/task-bird-audio-detection#audio-datasets\n\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019", "votes": 1 }, { "id": 895424, "postDate": "2020-06-21T10:35:58.537Z", "content": "torchvision models and pretrained weights\nhttps://pytorch.org/docs/stable/torchvision/models.html
", "rawMarkdown": "torchvision models and pretrained weights\nhttps://pytorch.org/docs/stable/torchvision/models.html", "votes": 1 }, { "id": 891063, "postDate": "2020-06-17T21:42:49.030Z", "content": "It appears that many competitors plan to train on external platforms (local or cloud). They then will upload their saved model_weights to kaggle for inclusion in their notebook. How is this covered by the External Data rules?
", "rawMarkdown": "It appears that many competitors plan to train on external platforms (local or cloud). They then will upload their saved model_weights to kaggle for inclusion in their notebook. How is this covered by the External Data rules?", "votes": 1, "replies": [ { "id": 891651, "postDate": "2020-06-18T11:06:37.420Z", "content": "I don't suppose weights constitute data.
\n\nAFAIU, you still need to abide by the external data rules, regardless where you train your models. Meaning, if you use external data / pretrained models as a starting point, you need to specify them in this thread.
\n\nShould you win the competition, you will have to share your code and demonstrate you followed the rules, so not sure there is any issue here with regards to where you chose to train your model.
\n\nWould be cool though if powers that may be confirmed my understanding 🙂
", "rawMarkdown": "I don't suppose weights constitute data.\n\nAFAIU, you still need to abide by the external data rules, regardless where you train your models. Meaning, if you use external data / pretrained models as a starting point, you need to specify them in this thread.\n\nShould you win the competition, you will have to share your code and demonstrate you followed the rules, so not sure there is any issue here with regards to where you chose to train your model. \n\nWould be cool though if powers that may be confirmed my understanding 🙂 ", "votes": 5 }, { "id": 893448, "postDate": "2020-06-19T16:00:36.863Z", "content": "Yes; we're running the competition to improve the science, so reproducibility is important - we really want to know what data was used, and how. So using external data is fine, just make sure you've done so in an ethical way, and disclose in the thread.
", "rawMarkdown": "Yes; we're running the competition to improve the science, so reproducibility is important - we really want to know what data was used, and how. So using external data is fine, just make sure you've done so in an ethical way, and disclose in the thread.", "votes": 2 } ] }, { "id": 889024, "postDate": "2020-06-16T18:18:24.860Z", "content": "The eBird Taxonomy: https://ebird.org/science/the-ebird-taxonomy
", "rawMarkdown": "The eBird Taxonomy: https://ebird.org/science/the-ebird-taxonomy", "votes": 1 }, { "id": 934686, "postDate": "2020-07-18T16:34:23.673Z", "content": "Additional recordings from https://www.xeno-canto.org/ as discussed here maintained as Kaggle dataset with appropriate per-record license:
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
https://github.com/timsainb/noisereduce
", "rawMarkdown": "https://github.com/timsainb/noisereduce" }, { "id": 1010895, "postDate": "2020-09-15T06:03:50.997Z", "content": "https://github.com/zhanghang1989/ResNeSt
\nhttps://pytorch.org/docs/stable/torchvision/models.html
pretrained weights from https://github.com/qubvel/classification_models
", "rawMarkdown": "pretrained weights from https://github.com/qubvel/classification_models" }, { "id": 1004928, "postDate": "2020-09-10T06:21:01.073Z", "content": "https://github.com/zhanghang1989/ResNeSt
\nhttps://github.com/lukemelas/EfficientNet-PyTorch
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttps://github.com/qiuqiangkong/audioset_tagging_cnn
https://github.com/zhanghang1989/ResNeSt
\nhttps://github.com/lukemelas/EfficientNet-PyTorch
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
https://github.com/UKYSpeechLab/ukybirddet
", "rawMarkdown": "https://github.com/UKYSpeechLab/ukybirddet" }, { "id": 1002688, "postDate": "2020-09-08T10:39:38.677Z", "content": "https://github.com/qiuqiangkong/audioset_tagging_cnn
\nhttps://pytorch.org/docs/stable/torchvision/models.html
\nhttps://github.com/Cadene/pretrained-models.pytorch
\nhttps://github.com/rwightman/pytorch-image-models
\nhttps://github.com/lukemelas/EfficientNet-PyTorch
\nhttps://github.com/zhanghang1989/ResNeSt
https://www.kaggle.com/mariotsaberlin/efficientnet-pytorch063
\nhttps://www.kaggle.com/mariotsaberlin/pytorch-resnest
Some audio files from the Animal Sound Archive Berlin (hosted by Museum für Naturkunde) containing species matching the training set:
\nhttps://www.kaggle.com/mariotsaberlin/animal-sound-archive-berlin
I did not use them yet but I am planning to add them to the training data.
", "rawMarkdown": "https://www.kaggle.com/mariotsaberlin/efficientnet-pytorch063\nhttps://www.kaggle.com/mariotsaberlin/pytorch-resnest\n\nSome audio files from the Animal Sound Archive Berlin (hosted by Museum für Naturkunde) containing species matching the training set:\nhttps://www.kaggle.com/mariotsaberlin/animal-sound-archive-berlin\n\nI did not use them yet but I am planning to add them to the training data." }, { "id": 1000466, "postDate": "2020-09-06T14:45:16.900Z", "content": "ResNeSt Pretrained Weights:
\nhttps://github.com/zhanghang1989/ResNeSt/
https://www.kaggle.com/vladimirsydor/resnest-git
\nhttps://www.kaggle.com/vladimirsydor/geffnet-pack
sorry not 100% sure how this works as not done many kaggle comps but think the things ive used are
\nresnet - imagenet weights
\nefficientnet - imagenet weights
\nkaggle freesound 2019 data - https://www.kaggle.com/c/freesound-audio-tagging-2019
\nenvironment sound data (also kaggle data set) - https://www.kaggle.com/mmoreaux/environmental-sound-classification-50
https://tf-explain.readthedocs.io/en/latest/
\n\n\nhttps://pjreddie.com/darknet/yolo/
\n\nhttps://pjreddie.com/darknet/imagenet/.
\nhttps://pjreddie.com/media/files/yolov3-tiny.weights
\nhttps://github.com/pythonlessons/TensorFlow-2.x-YOLOv3
\nhttps://pjreddie.com/media/files/yolov3.weights
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
https://www.kaggle.com/matrixneo/noise-birdsong-resampled-train-audio-00-ab
\nhttps://www.kaggle.com/matrixneo/noise-birdsong-resampled-train-audio-00-cf
\nhttps://www.kaggle.com/matrixneo/noise-birdsong-resampled-train-audio-00-gm
\nhttps://www.kaggle.com/matrixneo/noise-birdsong-resampled-train-audio-00-nr
\nhttps://www.kaggle.com/matrixneo/noise-birdsong-resampled-train-audio-00-sy
https://github.com/qiuqiangkong/torchlibrosa
", "rawMarkdown": "https://github.com/qiuqiangkong/torchlibrosa" }, { "id": 998689, "postDate": "2020-09-05T01:04:57.120Z", "content": "http://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\n https://www.kaggle.com/c/freesound-audio-tagging-2019/data
\nhttps://github.com/qubvel/efficientnet
\nhttp://dcase.community/challenge2018/task-bird-audio-detection#audio-datasets
https://www.kaggle.com/vladimirsydor/cornelli-background-noises
", "rawMarkdown": "https://www.kaggle.com/vladimirsydor/cornelli-background-noises", "replies": [ { "id": 1000290, "postDate": "2020-09-06T12:49:55.570Z", "content": "hi, what is the license for these files? where do they come from?
", "rawMarkdown": "hi, what is the license for these files? where do they come from?", "votes": 1 } ] }, { "id": 995810, "postDate": "2020-09-02T19:55:53.693Z", "content": "https://github.com/qiuqiangkong/audioset_tagging_cnn
", "rawMarkdown": "https://github.com/qiuqiangkong/audioset_tagging_cnn" }, { "id": 995605, "postDate": "2020-09-02T15:50:04.257Z", "content": "http://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
https://github.com/rwightman/gen-efficientnet-pytorch/tree/master/geffnet
", "rawMarkdown": "https://github.com/rwightman/gen-efficientnet-pytorch/tree/master/geffnet" }, { "id": 992186, "postDate": "2020-08-31T03:12:04.383Z", "content": "https://www.kaggle.com/samhiatt/xenocanto-avian-vocalizations-canv-usa
", "rawMarkdown": "https://www.kaggle.com/samhiatt/xenocanto-avian-vocalizations-canv-usa" }, { "id": 991062, "postDate": "2020-08-30T05:33:26.673Z", "content": "https://www.kaggle.com/luisblanche/birdcall-singing-0
\nhttps://www.kaggle.com/luisblanche/birdcall-singing-1
\nhttps://www.kaggle.com/luisblanche/birdcall-singing-2
\nhttps://www.kaggle.com/luisblanche/birdcall-singing-3
\nhttps://www.kaggle.com/luisblanche/birdcall-singing-4
\nhttps://www.kaggle.com/luisblanche/birdcall-background
By @luisblanche
\nhttps://www.kaggle.com/alanchn31/cornell-birdcall-nocalls
", "rawMarkdown": "https://www.kaggle.com/luisblanche/birdcall-singing-0\nhttps://www.kaggle.com/luisblanche/birdcall-singing-1\nhttps://www.kaggle.com/luisblanche/birdcall-singing-2\nhttps://www.kaggle.com/luisblanche/birdcall-singing-3\nhttps://www.kaggle.com/luisblanche/birdcall-singing-4\nhttps://www.kaggle.com/luisblanche/birdcall-background\n\nBy @luisblanche \n\nhttps://www.kaggle.com/alanchn31/cornell-birdcall-nocalls" }, { "id": 987804, "postDate": "2020-08-27T14:31:09.690Z", "content": "https://www.kaggle.com/mmoreaux/environmental-sound-classification-50
", "rawMarkdown": "https://www.kaggle.com/mmoreaux/environmental-sound-classification-50" }, { "id": 985125, "postDate": "2020-08-25T14:17:05.970Z", "content": "http://dcase.community/challenge2018/task-bird-audio-detection
\nhttps://github.com/karolpiczak/ESC-50
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://github.com/robbmcleod/pyfastnoisesimd
http://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://github.com/AgaMiko/bird-recognition-review
\nhttp://dcase.community/challenge2018/task-bird-audio-detection
\nhttps://github.com/qiuqiangkong/audioset_tagging_cnn
\nhttps://research.google.com/audioset
\nhttps://pytorch.org/docs/stable/torchvision/models.html
Ultralitics YOLOv5 with COCO pretrained weights:
\nhttps://github.com/ultralytics/yolov5
\nhttps://github.com/ultralytics/yolov5/releases/tag/v3.0
NIPS 2013 https://www.kaggle.com/c/multilabel-bird-species-classification-nips2013/data
", "rawMarkdown": "NIPS 2013 https://www.kaggle.com/c/multilabel-bird-species-classification-nips2013/data" }, { "id": 981639, "postDate": "2020-08-22T15:46:26.857Z", "content": "https://www.sciencedirect.com/science/article/pii/S1574954115000151 and http://taylor0.biology.ucla.edu/birdDBQuery/
", "rawMarkdown": "https://www.sciencedirect.com/science/article/pii/S1574954115000151 and http://taylor0.biology.ucla.edu/birdDBQuery/" }, { "id": 980252, "postDate": "2020-08-21T12:44:06.137Z", "content": "https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/SpeechSynthesis/Tacotron2
\nhttps://github.com/r9y9/wavenet_vocoder
\nhttps://keithito.com/LJ-Speech-Dataset/
\nhttp://www.festvox.org/cmu_arctic/
Pytorch Lightning\nhttps://www.kaggle.com/anjum48/pytorch-lightning
", "rawMarkdown": "Pytorch Lightning\n[https://www.kaggle.com/anjum48/pytorch-lightning](https://www.kaggle.com/anjum48/pytorch-lightning)" }, { "id": 964332, "postDate": "2020-08-09T18:44:03.263Z", "content": "", "rawMarkdown": "[Google Audioset](https://research.google.com/audioset/)" }, { "id": 963504, "postDate": "2020-08-09T04:47:28.327Z", "content": "Extended dataset as described here:
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
Keras-Applications\nhttps://files.pythonhosted.org/packages/71/e3/19762fdfc62877ae9102edf6342d71b28fbfd9dea3d2f96a882ce099b03f/Keras_Applications-1.0.8-py3-none-any.whl
", "rawMarkdown": "efficientnet\n[https://files.pythonhosted.org/packages/28/91/67848a143b54c331605bfba5fd31cf4e9db13d2e429d103fe807acc3bcf4/efficientnet-1.1.0-py3-none-any.whl](https://files.pythonhosted.org/packages/28/91/67848a143b54c331605bfba5fd31cf4e9db13d2e429d103fe807acc3bcf4/efficientnet-1.1.0-py3-none-any.whl)\n\nKeras-Applications\n[https://files.pythonhosted.org/packages/71/e3/19762fdfc62877ae9102edf6342d71b28fbfd9dea3d2f96a882ce099b03f/Keras_Applications-1.0.8-py3-none-any.whl](https://files.pythonhosted.org/packages/71/e3/19762fdfc62877ae9102edf6342d71b28fbfd9dea3d2f96a882ce099b03f/Keras_Applications-1.0.8-py3-none-any.whl)" }, { "id": 961822, "postDate": "2020-08-07T14:24:47.717Z", "content": "Audiomentations\nhttps://github.com/iver56/audiomentations
", "rawMarkdown": "Audiomentations\nhttps://github.com/iver56/audiomentations" }, { "id": 955643, "postDate": "2020-08-02T19:12:04.093Z", "content": "Preprocessed Spectrograms of bird calls\nhttps://www.kaggle.com/wabalubdubdub/bird-spectrogram-npz-a-to-z
\n\nSame bird calls as training set
", "rawMarkdown": "Preprocessed Spectrograms of bird calls\n[https://www.kaggle.com/wabalubdubdub/bird-spectrogram-npz-a-to-z]()\n\nSame bird calls as training set" }, { "id": 943261, "postDate": "2020-07-24T08:39:55.273Z", "content": "Microphone Impulse Response Project\nhttp://micirp.blogspot.com/
", "rawMarkdown": "Microphone Impulse Response Project\nhttp://micirp.blogspot.com/" }, { "id": 941952, "postDate": "2020-07-23T14:06:43.760Z", "content": "@ can I just mention https://tfhub.dev/ or do I have to share which specific model with weights from the tensorflow hub, I am using for my submission.
\n\nbtw, I am using models available at tensorflow hub at https://tfhub.dev/
", "rawMarkdown": "@ can I just mention https://tfhub.dev/ or do I have to share which specific model with weights from the tensorflow hub, I am using for my submission.\n\nbtw, I am using models available at tensorflow hub at https://tfhub.dev/" }, { "id": 937439, "postDate": "2020-07-21T02:53:41.160Z", "content": "@stefankahl - Is it ok to use sounds from freesound.org with license \"Attribution\" and \"CC0\" ? Do we have to list all audio files we use here or only \"freesound.org\" as a general source?
", "rawMarkdown": "@stefankahl - Is it ok to use sounds from freesound.org with license \"Attribution\" and \"CC0\" ? Do we have to list all audio files we use here or only \"freesound.org\" as a general source?", "replies": [ { "id": 943298, "postDate": "2020-07-24T09:09:33.180Z", "content": "I would say just listing freesound is ok. There are dedicated non-bird datasets however: http://machine-listening.eecs.qmul.ac.uk/bird-audio-detection-challenge/
", "rawMarkdown": "I would say just listing freesound is ok. There are dedicated non-bird datasets however: http://machine-listening.eecs.qmul.ac.uk/bird-audio-detection-challenge/ ", "votes": 1 }, { "id": 984014, "postDate": "2020-08-24T19:07:33.640Z", "content": "@stefankahl In addition to freefield1010 (the link related to freesound you provided above), can we also use the warblrb10k development dataset? It provides bird/nobird dataset.
\n8,000 smartphone audio recordings from around the UK, crowdsourced by users of Warblr the bird recognition app. The audio covers a wide distribution of UK locations and environments, and includes weather noise, traffic noise, human speech and even human bird imitations.
\nIt's under creative Commons Attribution licence CC-BY 4.0
\nhttp://dcase.community/challenge2018/task-bird-audio-detection
OK so: Freesound - pulled some non-bird sound samples from there.
\nIn addition I will use the extended train data that was approved in this discussion before.
", "rawMarkdown": "OK so: Freesound - pulled some non-bird sound samples from there.\n\nIn addition I will use the extended train data that was approved in this discussion before." } ] }, { "id": 936165, "postDate": "2020-07-20T03:19:00.653Z", "content": "ESC-50\nhttps://github.com/karolpiczak/ESC-50
", "rawMarkdown": "ESC-50\nhttps://github.com/karolpiczak/ESC-50" }, { "id": 932521, "postDate": "2020-07-17T05:49:43.770Z", "content": "fastai (pytorch) pretrained models: https://docs.fast.ai/vision.models.html
\nebird Regional range and abundance as tabular data: https://ebird.org/science/status-and-trends/download-data/download?package=all-stats-regional
\nplus codes: https://plus.codes/api
\nnoise samples: http://dcase.community/challenge2018/task-bird-audio-detection#audio-datasets
https://github.com/lukemelas/EfficientNet-PyTorch
\nhttps://github.com/huggingface/transformers
\nhttps://www.kaggle.com/ludovick/esc50datasetresampled
freefield1010, warblrb10k, BirdVox-DCASE-20k, Chernobyl, PolandNFC : http://dcase.community/challenge2018/task-bird-audio-detection
", "rawMarkdown": "https://github.com/lukemelas/EfficientNet-PyTorch\nhttps://github.com/huggingface/transformers\nhttps://www.kaggle.com/ludovick/esc50datasetresampled\n\nfreefield1010, warblrb10k, BirdVox-DCASE-20k, Chernobyl, PolandNFC : http://dcase.community/challenge2018/task-bird-audio-detection" }, { "id": 921806, "postDate": "2020-07-09T15:15:40.370Z", "content": "https://github.com/qubvel/efficientnet
\n\nhttps://github.com/keras-team/keras-applications
", "rawMarkdown": "https://github.com/qubvel/efficientnet\n\nhttps://github.com/keras-team/keras-applications\n" }, { "id": 893210, "postDate": "2020-06-19T13:09:54.657Z", "content": "Maybe I missed this in the rules but where is the data specified for posting links to external data sources?
", "rawMarkdown": "Maybe I missed this in the rules but where is the data specified for posting links to external data sources?" }, { "id": 1001853, "postDate": "2020-09-07T16:09:50.660Z", "rawMarkdown": "", "isDeleted": true }, { "id": 981986, "postDate": "2020-08-22T23:30:18.227Z", "rawMarkdown": "", "isDeleted": true }, { "id": 966429, "postDate": "2020-08-11T12:16:07.087Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 887971, "postDate": "2020-06-16T03:25:38.183Z", "rawMarkdown": "", "isDeleted": true }, { "id": 887550, "postDate": "2020-06-15T18:22:18.780Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 887538, "author_name": "robga", "author_url": "", "post_date": "2020-06-15T18:17:35.543000", "content": "Do we need additional permissions or licenses from individual birds appearing in any external dataset? (DFDC reference).
", "votes": 18, "replies": [] }, { "id": 911336, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-01T17:02:35.620000", "content": "In addition to my post earlier, here's the download link for the 2020 BirdCLEF validation data (which contains 60 minutes of soundscapes from North America and Peru): https://drive.google.com/file/d/1OIEJ9W6W5MNMXOicjiKivGmxTtskcJOn/view?usp=sharing
\n\nYou can use this data to test the performance of your system (of course, you'd have to adapt the ground truth to fit your needs). It is not permitted to use any other BirdCLEF data from current or past editions.
\n\nLet me know if you have any issues with the dataset.
", "votes": 14, "replies": [ { "id": 911528, "author_name": "CVxTz", "author_url": "", "post_date": "2020-07-01T19:57:44.560000", "content": "Is the training data of birdclef 2020 ok to use ? it is 70k clips from xeno canto.
", "votes": 0, "replies": [] }, { "id": 912040, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-02T07:09:33.317000", "content": "Well, not sure if it is going to be much of a help - it contains a lot of species that are irrelevant for this competition and it should contain the same recordings for species that are relevant for this competition. So I guess, yes, you can take a look but make sure to post it here if you find any portion of the data helpful.
", "votes": 0, "replies": [] }, { "id": 912709, "author_name": "Maxwell", "author_url": "", "post_date": "2020-07-02T17:04:59.243000", "content": ">it contains a lot of species that are irrelevant for this competition
\n\nYes... I confirmed type of species in BirdCLEF2020.\nPER recordings do not include any spices of this competition. Meanwhile, SSW recordings have 76 types of birds, but only 17 types of them match to this competition.\nI think if test datasets include all 264 species, BirdCLEF2020 dataset may not be sufficient to evaluate our models...
\n\n@stefankahl \nIf I'm wrong, please correct me.
\n\ncomments added,
\n\nAnd ground truth file is messy.\nSome rows are duplicated,\ne.g.\nSSW52_20170429 file has 3 duplicated rows.\nOne of them is like this,\n00:02:30-00:02:35 purfin\n00:02:30-00:02:35 purfin
", "votes": 2, "replies": [] }, { "id": 913586, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-03T09:44:49.703000", "content": "I manually looked at the BirdCLEF validation data and it seems that it contains labels for 16 species (I may have missed one) and all of them seem to be part of the training data for this competition. Not sure why the GT contains duplicates but it is save to ignore them. Sure, PER soundscapes are of no use but the SSW files should be a nice way to validate (yet, they are not the busiest soundscapes but suppressing false positives is key!).
\n\nAnd just as a reminder: We chose the training data according to lists of birds that MIGHT occur at the recording sites of the test data - not knowing which birds will actually vocalize is part of the challenge we're facing when deploying recorders.
", "votes": 6, "replies": [] }, { "id": 914431, "author_name": "CVxTz", "author_url": "", "post_date": "2020-07-03T21:32:26.383000", "content": "There are 76 labels in the validation set of Birdclef 2020, 17 of which are in this competition data.
", "votes": 0, "replies": [] }, { "id": 914782, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-04T08:17:20.203000", "content": "Well, I guess if you combine PER and SSW that's the case. But again, only the SSW soundscapes are of relevance for this competition, PER recordings are from South America - hence the lack of label overlap. All species annotated in the SSW files should be part of the training data for this competition and these soundscapes can be used for validation. We can't provide you with any other soundscape data since annotations are so hard to come by. Yet, optimizing the false positive rate of your classifier should be possible with the SSW soundscapes.
", "votes": 3, "replies": [] }, { "id": 914995, "author_name": "CVxTz", "author_url": "", "post_date": "2020-07-04T12:08:43.280000", "content": "Hello,
\n\nThanks for clarifying this, i missed the location information.
", "votes": 0, "replies": [] }, { "id": 925174, "author_name": "Shiro", "author_url": "", "post_date": "2020-07-11T21:31:14.430000", "content": "@stefankahl Thanks for the link. Can we use this data for the training part ?
", "votes": 0, "replies": [] }, { "id": 959531, "author_name": "", "author_url": "", "post_date": "2020-08-05T17:04:14.193000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 893651, "author_name": "Vopani", "author_url": "", "post_date": "2020-06-19T19:10:20.760000", "content": "Additional recordings from https://www.xeno-canto.org/
\nSome are discussed here and maintained as Kaggle datasets with appropriate per-record license:
https://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition
\nhttps://github.com/qiuqiangkong/audioset_tagging_cnn
ResNeSt pre-trained weights using in my baseline
\nhttps://github.com/zhanghang1989/ResNeSt
pytorch-pfn-extras (for training)
\nhttps://github.com/pfnet/pytorch-pfn-extras
previous kaggle challenges\n* https://www.kaggle.com/c/mlsp-2013-birds/data\n* https://www.kaggle.com/c/multilabel-bird-species-classification-nips2013/data
", "votes": 6, "replies": [ { "id": 903610, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2020-06-27T01:45:32.180000", "content": "niiiiiiiiice
", "votes": 0, "replies": [] } ] }, { "id": 914485, "author_name": "Tanrei(nama)", "author_url": "", "post_date": "2020-07-04T00:47:06.470000", "content": "Environmental sound and noise sound for nocall.\nhttps://www.youtube.com/watch?v=xNN7iTA57jM\nhttps://www.youtube.com/watch?v=8plwv25NYRo\nhttps://www.youtube.com/watch?v=lR4GNWcwAI8\nhttps://www.youtube.com/watch?v=4KzFe50RQkQ
", "votes": 3, "replies": [ { "id": 996135, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2020-09-03T05:37:23.767000", "content": "haha dude that audio is FULL of birdcalls!
", "votes": 0, "replies": [] } ] }, { "id": 953117, "author_name": "beluga", "author_url": "", "post_date": "2020-07-31T14:44:48.177000", "content": "[Animal Sound Archive] ()https://www.gbif.org/dataset/b7ec1bf8-819b-11e2-bad2-00145eb45e9a) Published by Museum für Naturkunde Berlin
\n\nThe Animal Sound Archive at the Museum fuer Naturkunde Berlin (German: Tierstimmenarchiv) is one of the oldest and largest worldwide. Founded in 1951 by Professor Guenter Tembrock the collection consists now of around 130 000 records of animal voices.
", "votes": 4, "replies": [] }, { "id": 943332, "author_name": "marshi", "author_url": "", "post_date": "2020-07-24T09:52:22.280000", "content": "Microphone wind noise simulator\nhttps://github.com/kenders2000/MicWindNoiseGenerator
", "votes": 4, "replies": [] }, { "id": 935200, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2020-07-19T06:30:17.233000", "content": "I just learned about this very interesting resource - radio aporee. It is an enormous collection of soundscape recordings from all over the world! 🙂 Might be very useful to see how our models perform on data coming from various devices.
", "votes": 4, "replies": [] }, { "id": 898238, "author_name": "beluga", "author_url": "", "post_date": "2020-06-23T11:49:58.650000", "content": "I was looking for soundscape recordings and found the LifeCLEF Bird challenges.
\n\n\n\n\nThe test data consists of 153 soundscapes recorded in Peru, the USA, and Germany. Each soundscape is of ten-minute duration and contains high quantities of (overlapping) bird vocalizations.
\n
I am bit worried as the two test sets might overlap.
\n\n\n", "votes": 4, "replies": [ { "id": 898243, "author_name": "Vopani", "author_url": "", "post_date": "2020-06-23T11:53:32", "content": "The hidden test_audio directory contains approximately 150 recordings in mp3 format, each roughly 10 minutes long.
\n
I see you are trying hard to find the test data 😁 \nBut this test data looks eerily similar.
", "votes": 1, "replies": [] }, { "id": 898253, "author_name": "beluga", "author_url": "", "post_date": "2020-06-23T11:58:01.557000", "content": "I found the CLEF competitions by clicking twice starting from this discussion thread :)
\n\nSince I found them it is better to share early.\nAnyway I would like to create useful local validation set with soundscape recordings. I don't like when local CV does not match LB...
", "votes": 4, "replies": [] }, { "id": 898655, "author_name": "Sohier Dane", "author_url": "", "post_date": "2020-06-23T16:54:12.127000", "content": "The LifeClef files appear similar because some of the same organizations were involved in generating the datasets used by each competition. However, there is no overlap between the files.
", "votes": 6, "replies": [] }, { "id": 898667, "author_name": "beluga", "author_url": "", "post_date": "2020-06-23T17:01:38.913000", "content": "So we are allowed to use them (even manually annotate & train models on them), right?
", "votes": 1, "replies": [] }, { "id": 899494, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-06-24T09:18:10.263000", "content": "Here's my concern: The LifeCLEF dataset is not explicitly public - you have to register to get the files and they will be gone once the second round of submissions is over. We have to ensure that this actually qualifies as a valid external source according to the rules. We will look into that and keep you posted.
", "votes": 0, "replies": [] }, { "id": 899523, "author_name": "beluga", "author_url": "", "post_date": "2020-06-24T09:36:11.487000", "content": "@stefankahl thanks for looking into that, I saw you are also an organizer in those competitions. Btw I could register and download the closed (2018) competition data as well. If we are allowed I am happy to publish the datasets here at kaggle although could not find any licensing terms in the rules.
", "votes": 1, "replies": [] }, { "id": 899696, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-06-24T11:53:31.823000", "content": "The 2018 data does not contain any North American soundscapes (and no labels). The 2020 dataset would be more interesting, although I think we should avoid to allow training on soundscape data (even though there's no overlap with the Kaggle test set, it just wouldn't fit the task). But it would be a nice validation dataset - we will discuss that and let you know.
", "votes": 5, "replies": [] }, { "id": 902798, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2020-06-26T11:16:21.950000", "content": "Having a validation set could be really useful for figuring out how to go from training a classifier on train data to predicting on soundscape. Right now we are essentially flying blind in this regard.
\n\nI am also curious, @stefankahl would you by any chance be aware of a resource that would list birds along with how common they are to certain regions, whether they are transitory, etc? Would be ideal if this contained an ebird code but if not that is not a problem. I looked on ebird.org but maybe I was looking in the wrong place or this information does not exist there.
\n\nThanks a lot for all your help!
", "votes": 3, "replies": [] }, { "id": 903603, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2020-06-27T01:27:55.200000", "content": "@stefankahl I think some kind of soundscape data including multiple birdcalls would confer a huge benefit to this competition and improve the resulting models since it would allow participants to estimate the strength of their models (and certain important hyperparameters like prediction score threshold) WITHOUT mercilessly probing the public test set.
\n\nJust my two cents.
", "votes": 4, "replies": [] }, { "id": 903607, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2020-06-27T01:37:13.487000", "content": "\n\n\nThe 2018 data does not contain any North American soundscapes
\n
ha-HA! so all 3 sites are in North America! We're on to you Stefan.
", "votes": 0, "replies": [] }, { "id": 911073, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-01T14:53:09.127000", "content": "Hm, not sure if this was a secret. The training data only contains North American species :)
", "votes": 2, "replies": [] }, { "id": 911089, "author_name": "beluga", "author_url": "", "post_date": "2020-07-01T15:03:30.330000", "content": "Maybe it is worth mentioning on the overview/data page. I only found this info in the forum.
", "votes": 1, "replies": [] }, { "id": 911091, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-01T15:04:02.683000", "content": "I talked to the other hosts about the use of BirdCLEF data for validation and we decided that this is not going to be problematic. Therefore, I put the 2020 validation data on Google Drive so you don't have to register for BirdCLEF (unless, of course, you like to submit, the challenge is still open for post-deadline submissions). In any case, the dataset contains 60 minutes of soundscapes from Peru (probably not of much use for this competition) and 60 minutes from North America (no overlap with the Kaggle test data, but as you mentioned it could be valuable to test your systems). I think the Kaggle training data should cover all species from the BirdCLEF validation data, but the contained soundscapes aren't the busiest ones. I know I can't prevent you from training on soundscape data, but I would like to remind you of the goal of this competition: Train on focal recordings and apply to soundscapes - that's the only way that we can apply a detection system to a new location without annotating hours of soundscapes first. Due to this, the use of other data from BirdCLEF will not be permitted (please do not annotate other soundscapes, that would be against the competition's intent).
\n\nHere's the download link: https://drive.google.com/file/d/1OIEJ9W6W5MNMXOicjiKivGmxTtskcJOn/view?usp=sharing
\n\nLet me know if you have any issues with the dataset.
", "votes": 8, "replies": [] }, { "id": 911120, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-01T15:21:22.170000", "content": "Regarding good resources on bird species: I think eBird.org is a good starting point. They have range maps for all species and so-called \"bar charts\" where you can see the levels of abundance for a given location. (e.g., for the American Robin you can browse to https://ebird.org/species/amerob). But I also think that the recording locations and dates in the training data should reflect species abundance across all seasons and you should be able to figure out if a species is migrating or not without manual interference. The number of recordings (in the extended training data) and diversity of recording locations should also reflect how common a species is.
\n\nBTW: eBird also has a nice public API (https://documenter.getpostman.com/view/664302/S1ENwy59?version=latest), yet I am not sure if any of the features it offers would be suited to help you for this competition.
\n\nAnd again, as a reminder: Please do not crawl these sites without considering the load you might impose on the servers.
", "votes": 3, "replies": [] }, { "id": 911171, "author_name": "Vopani", "author_url": "", "post_date": "2020-07-01T15:44:33.277000", "content": "@stefankahl Thanks a lot for the updates and for sharing the data. Though I strongly recommended to post these comments / data as a separate pinned topic (like the XC-crawling one) or part of the official rules / official data, so that it is clearly visible to all competitors (rather than lost in the middle of this thread) especially considering some of these points affect the rules of competing / modelling in this competition.\ncc: @sohier
", "votes": 1, "replies": [] }, { "id": 911342, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-01T17:06:33.617000", "content": "I added another post that contains the link. However, I think this is the right place to announce it since it is additional data which is not part of the official dataset (despite being announced by a host). But of course, feel free to post another thread or notebook that explores the recordings and maybe even manages to provide a ground truth that fits this competition (right now, it's BirdCLEF-specific).
", "votes": 4, "replies": [] }, { "id": 911641, "author_name": "", "author_url": "", "post_date": "2020-07-01T22:37:10.527000", "content": "", "votes": 0, "replies": [] }, { "id": 916583, "author_name": "", "author_url": "", "post_date": "2020-07-05T19:46:33.150000", "content": "", "votes": 0, "replies": [] }, { "id": 963132, "author_name": "MPWARE", "author_url": "", "post_date": "2020-08-08T17:20:19.673000", "content": "@stefankahl You said test set contains only North America species correct? If I'm not wrong there are 262 species over 264 covering North America in train set. There are 2 species never seen in North America:\n- gnwtea: Eurasian Teal\n- hergul: European Herring Gull
\n\nIs that correct? If so, we can label them as \"nocall\" on inference time if there are detected.\nIt won't make a big difference but it's good to know.
", "votes": 1, "replies": [] }, { "id": 965399, "author_name": "Sohier Dane", "author_url": "", "post_date": "2020-08-10T15:46:51.127000", "content": "The test set recording locations are in America. Birds occasionally take a wrong turn during a migration or get blown about in a storm and end up on the wrong continent so I would guess that the recordings of those two species in the train set were of rare sightings in North America. Birds seen outside of their usual territories are very popular with birders, it's not too surprising that that they would be recorded.
", "votes": 3, "replies": [] }, { "id": 965493, "author_name": "Tom Denton", "author_url": "", "post_date": "2020-08-10T17:10:25.847000", "content": "Couple things could be going on here... a) We're using the eBird codes, rather than the 6-letter codes. They MOSTLY agree, except for where they don't. b) Both of these birds have American and Eurasian flavors.
\n\ngnwtea == Green Winged Teal\nhergul == Herring Gull
", "votes": 2, "replies": [] }, { "id": 966159, "author_name": "something4kag", "author_url": "", "post_date": "2020-08-11T07:41:44.830000", "content": "wrt gnwtea and hergul, I did some checking on the location info in train and none were from North America. however, figured it was best to retain them just in case of migration timing not represented in train or other bird misadventures.
\n\nfor external data would like to add \nhttps://github.com/rwightman/pytorch-image-models
", "votes": 0, "replies": [] } ] }, { "id": 1006880, "author_name": "Balraj Ashwath", "author_url": "", "post_date": "2020-09-11T15:58:05.493000", "content": "https://www.kaggle.com/ludovick/xenoexternalwav0
\nhttps://www.kaggle.com/ludovick/xenoexternalwav1
EfficientNet-PyTorch:https://github.com/lukemelas/EfficientNet-PyTorch
\nPANN https://github.com/qiuqiangkong/audioset_tagging_cnn/
\nGoogle AudioSet https://research.google.com/audioset/download.html
\nfastai and pre-trained models: https://github.com/fastai/fastai
\nPretrained models for Pytorch https://github.com/cadene/pretrained-models.pytorch
\ntorchvision models and pretrained weights https://pytorch.org/docs/stable/torchvision/models.html
\nESC-50 https://github.com/karolpiczak/ESC-50
\nkeras https://keras.io/
\nkeras pre-trained models https://keras.io/api/applications/
Some human speech and general sounds for NN pretraining:
\nhttps://commonvoice.mozilla.org/en/datasets
\nhttps://www.openslr.org/17/
\nhttps://www.openslr.org/28/
Are optimizers and schedulers part of the External Data Thread and need to be reported? Have not seen that kind of post before in any other competition, like for AdamW,Radam,schedulers,dataloaders that are not built-in, so it's good to know.
", "votes": 1, "replies": [ { "id": 990406, "author_name": "leodav", "author_url": "", "post_date": "2020-08-29T15:26:06.523000", "content": "I am not sure if organizers though of all these details… It would be great if kaggle established a few version of rules that competons had to choose one. Think competitionright like copyright and authorship right.
", "votes": 0, "replies": [] }, { "id": 995860, "author_name": "Kirderf", "author_url": "", "post_date": "2020-09-02T21:52:55.320000", "content": "This is my second competition with exactly the same question in the same topic, did not get an answer then but did not repeat the question, but now I start to wonder, should I tag one from the team to get an answer or how does it work.
", "votes": 0, "replies": [] }, { "id": 998011, "author_name": "Kirderf", "author_url": "", "post_date": "2020-09-04T12:21:00.967000", "content": "@addisonhoward grateful for an answer when time allows.
", "votes": 0, "replies": [] }, { "id": 1003058, "author_name": "Addison Howard", "author_url": "", "post_date": "2020-09-08T16:23:03.687000", "content": "The competition hosts are ultimately the ones who are responsible for the competition rules, however the external data rule is designed, in spirit, to prevent someone with a \"silver bullet\" dataset to end up winning simply because they had access that nobody else did. Optimizers and schedulers aren't a separate dataset, but are a part of your workflow, and do not require disclosure.
", "votes": 1, "replies": [] } ] }, { "id": 979621, "author_name": "林湧森 (Dyson Lin)", "author_url": "", "post_date": "2020-08-21T02:04:12.503000", "content": "jhartquist's fastai_audio:
\nhttps://github.com/jhartquist/fastai_audio
mogwai's fastai_audio:
\nhttps://github.com/mogwai/fastai_audio
fire:
\nhttps://pypi.org/project/fire/
fastai2_audio:
\nhttps://github.com/rbracco/fastai2_audio
EfficientNet-PyTorch:
\nhttps://github.com/lukemelas/EfficientNet-PyTorch
fastai:
\nhttps://github.com/fastai/fastai
fastbook:
\nhttps://pypi.org/project/fastbook/
resnet18:
\nhttps://download.pytorch.org/models/resnet18-5c106cde.pth
kapre:
\nhttps://github.com/keunwoochoi/kapre
XC463492 · 大山雀 · Parus major:
\nhttps://www.xeno-canto.org/463492
XC464650 · 大山雀 · Parus major:
\nhttps://www.xeno-canto.org/464650
Can I use data from this challenge? https://www.kaggle.com/c/freesound-audio-tagging-2019/data
", "votes": 1, "replies": [] }, { "id": 968344, "author_name": "leodav", "author_url": "", "post_date": "2020-08-12T22:57:03.753000", "content": "https://www.kaggle.com/anjum48/pytorch-lightning
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttp://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m
\nhttps://github.com/qiuqiangkong/audioset_tagging_cnn
\nhttps://pytorch.org/docs/stable/torchvision/models.html
Environmental Sound Classification 50
", "votes": 1, "replies": [] }, { "id": 946130, "author_name": "yuvaramsingh", "author_url": "", "post_date": "2020-07-26T11:27:00.380000", "content": "Pls confirm if BirdVox data can be used for this competition
\n\nhttps://wp.nyu.edu/birdvox/birdvox-full-night/
", "votes": 1, "replies": [ { "id": 946165, "author_name": "Stefan Kahl", "author_url": "", "post_date": "2020-07-26T12:04:33.597000", "content": "BirdVox data can be used, but be aware that night flight calls are often fundamentally different from the types of vocalizations recorded in the train and test data.
", "votes": 5, "replies": [] } ] }, { "id": 940271, "author_name": "hengck23", "author_url": "", "post_date": "2020-07-22T19:40:50.247000", "content": "https://github.com/AgaMiko/bird-recognition-review\nhttp://dcase.community/challenge2018/task-bird-audio-detection
", "votes": 1, "replies": [] }, { "id": 918196, "author_name": "Daizu", "author_url": "", "post_date": "2020-07-07T04:26:19.620000", "content": "Pretrained models for Pytorch\nhttps://github.com/cadene/pretrained-models.pytorch
", "votes": 1, "replies": [] }, { "id": 909771, "author_name": "beluga", "author_url": "", "post_date": "2020-06-30T20:13:06.150000", "content": "", "votes": 1, "replies": [] }, { "id": 895902, "author_name": "CVxTz", "author_url": "", "post_date": "2020-06-21T17:16:19.917000", "content": "All 3 training and 3 validation datasets for this challenge.\nhttp://dcase.community/challenge2018/task-bird-audio-detection#audio-datasets
\n\nhttps://www.kaggle.com/c/freesound-audio-tagging-2019
", "votes": 1, "replies": [] }, { "id": 895424, "author_name": "Hidehisa Arai", "author_url": "", "post_date": "2020-06-21T10:35:58.537000", "content": "torchvision models and pretrained weights\nhttps://pytorch.org/docs/stable/torchvision/models.html
", "votes": 1, "replies": [] }, { "id": 891063, "author_name": "David Patton", "author_url": "", "post_date": "2020-06-17T21:42:49.030000", "content": "It appears that many competitors plan to train on external platforms (local or cloud). They then will upload their saved model_weights to kaggle for inclusion in their notebook. How is this covered by the External Data rules?
", "votes": 1, "replies": [ { "id": 891651, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2020-06-18T11:06:37.420000", "content": "I don't suppose weights constitute data.
\n\nAFAIU, you still need to abide by the external data rules, regardless where you train your models. Meaning, if you use external data / pretrained models as a starting point, you need to specify them in this thread.
\n\nShould you win the competition, you will have to share your code and demonstrate you followed the rules, so not sure there is any issue here with regards to where you chose to train your model.
\n\nWould be cool though if powers that may be confirmed my understanding 🙂
", "votes": 5, "replies": [] }, { "id": 893448, "author_name": "Tom Denton", "author_url": "", "post_date": "2020-06-19T16:00:36.863000", "content": "Yes; we're running the competition to improve the science, so reproducibility is important - we really want to know what data was used, and how. So using external data is fine, just make sure you've done so in an ethical way, and disclose in the thread.
", "votes": 2, "replies": [] } ] }, { "id": 889024, "author_name": "David Patton", "author_url": "", "post_date": "2020-06-16T18:18:24.860000", "content": "The eBird Taxonomy: https://ebird.org/science/the-ebird-taxonomy
", "votes": 1, "replies": [] }, { "id": 934686, "author_name": "Jatin Patni", "author_url": "", "post_date": "2020-07-18T16:34:23.673000", "content": "Additional recordings from https://www.xeno-canto.org/ as discussed here maintained as Kaggle dataset with appropriate per-record license:
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-n-z
\nhttps://www.kaggle.com/rohanrao/xeno-canto-bird-recordings-extended-a-m