So we can ignore the sample of the classes of the metadata that are not in the train_soundscapes.
", "messages": [ { "id": "1295050", "postDate": "05/06/2021 07:00:09", "content": "So we can ignore the sample of the classes of the metadata that are not in the train_soundscapes.
", "rawMarkdown": "So we can ignore the sample of the classes of the metadata that are not in the train_soundscapes.", "votes": null }, { "id": "1295144", "postDate": "05/06/2021 08:35:03", "content": "test set includes locations not present in train data hence restricting species may be a risky bet.
", "rawMarkdown": "test set includes locations not present in train data hence restricting species may be a risky bet.", "votes": null }, { "id": "1295170", "postDate": "05/06/2021 08:57:45", "content": "The metadata has 398 classes while the train_soundscapes only contains 49 classes. If the test set also has 49 classes then I may want to ignore the rest.
", "rawMarkdown": "The metadata has 398 classes while the train_soundscapes only contains 49 classes. If the test set also has 49 classes then I may want to ignore the rest.", "votes": null }, { "id": "1305565", "postDate": "05/13/2021 11:15:50", "content": "As a starter this is an interesting question to me. Is there an answer to it? Meaning, does the test data only contain classes from train_soundscapes or also some that are not present in train_soundscapes?
My bet is: train_soundscape < test_soundscape < metadata/snippets
", "rawMarkdown": "My bet is: train_soundscape < test_soundscape < metadata/snippets", "votes": null }, { "id": "1305827", "postDate": "05/13/2021 13:50:19", "content": "I agree with @ilu000.
\nSome species can be eliminated based on geographic range. Definitely no need to bother with Carolina Chickadee (carchi), for example; it is exceptionally unlikely to occur at any of the four test sites, as can be seen on the eBird range map:
\nhttps://ebird.org/species/carchi
Of the 398 species with short training audio, my estimates are that perhaps 315 of them are likely to occur at one or more of the four test sites at the time of year of the test soundscapes. But there's definitely slush in that estimate, as it is based on some assumptions that may or may not be accurate.
", "rawMarkdown": "I agree with @ilu000.\n\nSome species can be eliminated based on geographic range. Definitely no need to bother with Carolina Chickadee (carchi), for example; it is exceptionally unlikely to occur at any of the four test sites, as can be seen on the eBird range map:\nhttps://ebird.org/species/carchi\n\nOf the 398 species with short training audio, my estimates are that perhaps 315 of them are likely to occur at one or more of the four test sites at the time of year of the test soundscapes. But there's definitely slush in that estimate, as it is based on some assumptions that may or may not be accurate.", "votes": null }, { "id": "1306180", "postDate": "05/13/2021 16:49:08", "content": "The train soundscapes don't cover all of the recording locations, but all species in the test set are present in the short train audio clips.
", "rawMarkdown": "The train soundscapes don't cover all of the recording locations, but all species in the test set are present in the short train audio clips.", "votes": null }, { "id": "1306322", "postDate": "05/13/2021 17:44:47", "content": "@sohier Thanks. I made the same answer, but coming from you it may be trusted more ;)
", "rawMarkdown": "sohier Thanks. I made the same answer, but coming from you it may be trusted more ;)", "votes": null }, { "id": "1306392", "postDate": "05/13/2021 18:25:51", "content": "Complete test set may have more sites. I wouldn't rule out any specific species. But there may for example be very rare ones with only single occurence or even zero occurence.
\nEdit: seeing the comment below, seems I really need to read up a bit.
", "rawMarkdown": "Complete test set may have more sites. I wouldn't rule out any specific species. But there may for example be very rare ones with only single occurence or even zero occurence. \n\nEdit: seeing the comment below, seems I really need to read up a bit.", "votes": null }, { "id": "1306583", "postDate": "05/13/2021 22:24:09", "content": "The complete test set has four sites; we've been given the approximate locations (lat/lon to two decimal points) as well as the recording dates (which I assume is a complete list?). Combining that information with data from eBird, there are some species that have a near-zero probability of being present or heard at any of the four sites.
\nThe Carolina Chickadee is one that I have confidence as a birder will not be in any of the recordings for the four sites listed. Sapsucker Woods (New York) is the only test location remotely close to the Carolina Chickadee range, but Sapsucker Woods is firmly in Black-capped Chickadee territory. An unusually sharp boundary exists between the ranges for these two closely related species. Sapsucker Woods (being the home of the Cornell Lab of Ornithology) is ultra-heavily documented on eBird, yet a Carolina Chickadee has never been reported in Sapsucker Woods or its county, Tompkins County, New York. eBird does capture vagrants, and Tompkins County has some spectacular ones (Magnificent Frigatebird, for example), but literally no Carolina Chickadees.
\nThe Carolina Chickadee is a case of a species that can be omitted with no harm in the context of the four test sites. I happen to think there are a number of others, though there’s a gray area of determining just how unlikely of a species to include. The fact that eBird coverage and effort isn’t equivalent for all four test regions should also be considered.
", "rawMarkdown": "The complete test set has four sites; we've been given the approximate locations (lat/lon to two decimal points) as well as the recording dates (which I assume is a complete list?). Combining that information with data from eBird, there are some species that have a near-zero probability of being present or heard at any of the four sites. \n\nThe Carolina Chickadee is one that I have confidence as a birder will not be in any of the recordings for the four sites listed. Sapsucker Woods (New York) is the only test location remotely close to the Carolina Chickadee range, but Sapsucker Woods is firmly in Black-capped Chickadee territory. An unusually sharp boundary exists between the ranges for these two closely related species. Sapsucker Woods (being the home of the Cornell Lab of Ornithology) is ultra-heavily documented on eBird, yet a Carolina Chickadee has never been reported in Sapsucker Woods or its county, Tompkins County, New York. eBird does capture vagrants, and Tompkins County has some spectacular ones (Magnificent Frigatebird, for example), but literally no Carolina Chickadees.\n\nThe Carolina Chickadee is a case of a species that can be omitted with no harm in the context of the four test sites. I happen to think there are a number of others, though there’s a gray area of determining just how unlikely of a species to include. The fact that eBird coverage and effort isn’t equivalent for all four test regions should also be considered.", "votes": null }, { "id": "1308108", "postDate": "05/15/2021 02:34:08", "content": "@sohier Thanks for the answer. Does test set include species present only in primary labels in short train audio clips or those in secondary labels as well. Thanks!
", "rawMarkdown": "sohier Thanks for the answer. Does test set include species present only in primary labels in short train audio clips or those in secondary labels as well. Thanks!", "votes": null }, { "id": "1308803", "postDate": "05/15/2021 13:09:20", "content": "From Stefan Kahl's BirdCLEF2021: Exploring the data:
\n\n\nPlease note, secondary lables only contain labels of species that are actually represented in the data set.
\n
I take that to mean that the short train audio clips do not list any secondary species that aren't already included in the primary list (though I technically haven't formally checked this).
\nI also infer this to mean that other species might be in the background of the short train audio clips, but they wouldn't be listed. As Stefan also notes, \"lists of secondary labels are not very reliable\".
", "rawMarkdown": "From Stefan Kahl's [BirdCLEF2021: Exploring the data](https://www.kaggle.com/stefankahl/birdclef2021-exploring-the-data):\n\n> Please note, secondary lables only contain labels of species that are actually represented in the data set.\n\nI take that to mean that the short train audio clips do not list any secondary species that aren't already included in the primary list (though I technically haven't formally checked this). \n\nI also infer this to mean that other species might be in the background of the short train audio clips, but they wouldn't be listed. As Stefan also notes, \"lists of secondary labels are not very reliable\".", "votes": null }, { "id": "1308822", "postDate": "05/15/2021 13:23:43", "content": "@jmreuter I think some scondary label species are present in primary labels, while some don't.
", "rawMarkdown": "jmreuter I think some scondary label species are present in primary labels, while some don't.", "votes": null }, { "id": "1309049", "postDate": "05/15/2021 16:14:56", "content": "@tonychenxyz I just remembered that the primary vs secondary label question has also been discussed here, with answers from Stefan Kahl:
\nhttps://www.kaggle.com/c/birdclef-2021/discussion/231106
Short summary: If you treat rocpig and rocpig1 as one class (in the secondary labels, both representing the Rock Pigeon), then there are no secondary labels that are not also in the primary label set.
", "rawMarkdown": "tonychenxyz I just remembered that the primary vs secondary label question has also been discussed here, with answers from Stefan Kahl:\nhttps://www.kaggle.com/c/birdclef-2021/discussion/231106\n\nShort summary: If you treat rocpig and rocpig1 as one class (in the secondary labels, both representing the Rock Pigeon), then there are no secondary labels that are not also in the primary label set.", "votes": null }, { "id": "1309547", "postDate": "05/16/2021 05:01:12", "content": "@jmreuter This is interesting. Thanks for the info!
", "rawMarkdown": "jmreuter This is interesting. Thanks for the info!", "votes": null } ], "comments": [ { "id": 1295144, "author_name": "cpmpml", "author_url": "", "post_date": "05/06/2021 08:35:03", "content": "test set includes locations not present in train data hence restricting species may be a risky bet.
", "votes": null, "replies": [ { "id": 1295170, "author_name": "snaker", "author_url": "", "post_date": "05/06/2021 08:57:45", "content": "The metadata has 398 classes while the train_soundscapes only contains 49 classes. If the test set also has 49 classes then I may want to ignore the rest.
", "votes": null, "replies": [] }, { "id": 1305575, "author_name": "ilu000", "author_url": "", "post_date": "05/13/2021 11:23:41", "content": "My bet is: train_soundscape < test_soundscape < metadata/snippets
", "votes": null, "replies": [] }, { "id": 1305827, "author_name": "jmreuter", "author_url": "", "post_date": "05/13/2021 13:50:19", "content": "I agree with @ilu000.
\nSome species can be eliminated based on geographic range. Definitely no need to bother with Carolina Chickadee (carchi), for example; it is exceptionally unlikely to occur at any of the four test sites, as can be seen on the eBird range map:
\nhttps://ebird.org/species/carchi
Of the 398 species with short training audio, my estimates are that perhaps 315 of them are likely to occur at one or more of the four test sites at the time of year of the test soundscapes. But there's definitely slush in that estimate, as it is based on some assumptions that may or may not be accurate.
", "votes": null, "replies": [] }, { "id": 1306392, "author_name": "ilu000", "author_url": "", "post_date": "05/13/2021 18:25:51", "content": "Complete test set may have more sites. I wouldn't rule out any specific species. But there may for example be very rare ones with only single occurence or even zero occurence.
\nEdit: seeing the comment below, seems I really need to read up a bit.
", "votes": null, "replies": [] }, { "id": 1306583, "author_name": "jmreuter", "author_url": "", "post_date": "05/13/2021 22:24:09", "content": "The complete test set has four sites; we've been given the approximate locations (lat/lon to two decimal points) as well as the recording dates (which I assume is a complete list?). Combining that information with data from eBird, there are some species that have a near-zero probability of being present or heard at any of the four sites.
\nThe Carolina Chickadee is one that I have confidence as a birder will not be in any of the recordings for the four sites listed. Sapsucker Woods (New York) is the only test location remotely close to the Carolina Chickadee range, but Sapsucker Woods is firmly in Black-capped Chickadee territory. An unusually sharp boundary exists between the ranges for these two closely related species. Sapsucker Woods (being the home of the Cornell Lab of Ornithology) is ultra-heavily documented on eBird, yet a Carolina Chickadee has never been reported in Sapsucker Woods or its county, Tompkins County, New York. eBird does capture vagrants, and Tompkins County has some spectacular ones (Magnificent Frigatebird, for example), but literally no Carolina Chickadees.
\nThe Carolina Chickadee is a case of a species that can be omitted with no harm in the context of the four test sites. I happen to think there are a number of others, though there’s a gray area of determining just how unlikely of a species to include. The fact that eBird coverage and effort isn’t equivalent for all four test regions should also be considered.
", "votes": null, "replies": [] } ] }, { "id": 1305565, "author_name": "philippsinger", "author_url": "", "post_date": "05/13/2021 11:15:50", "content": "As a starter this is an interesting question to me. Is there an answer to it? Meaning, does the test data only contain classes from train_soundscapes or also some that are not present in train_soundscapes?
The train soundscapes don't cover all of the recording locations, but all species in the test set are present in the short train audio clips.
", "votes": null, "replies": [] }, { "id": 1306322, "author_name": "cpmpml", "author_url": "", "post_date": "05/13/2021 17:44:47", "content": "@sohier Thanks. I made the same answer, but coming from you it may be trusted more ;)
", "votes": null, "replies": [] }, { "id": 1308108, "author_name": "tonychenxyz", "author_url": "", "post_date": "05/15/2021 02:34:08", "content": "@sohier Thanks for the answer. Does test set include species present only in primary labels in short train audio clips or those in secondary labels as well. Thanks!
", "votes": null, "replies": [] }, { "id": 1308803, "author_name": "jmreuter", "author_url": "", "post_date": "05/15/2021 13:09:20", "content": "From Stefan Kahl's BirdCLEF2021: Exploring the data:
\n\n\nPlease note, secondary lables only contain labels of species that are actually represented in the data set.
\n
I take that to mean that the short train audio clips do not list any secondary species that aren't already included in the primary list (though I technically haven't formally checked this).
\nI also infer this to mean that other species might be in the background of the short train audio clips, but they wouldn't be listed. As Stefan also notes, \"lists of secondary labels are not very reliable\".
", "votes": null, "replies": [] }, { "id": 1308822, "author_name": "tonychenxyz", "author_url": "", "post_date": "05/15/2021 13:23:43", "content": "@jmreuter I think some scondary label species are present in primary labels, while some don't.
", "votes": null, "replies": [] }, { "id": 1309049, "author_name": "jmreuter", "author_url": "", "post_date": "05/15/2021 16:14:56", "content": "@tonychenxyz I just remembered that the primary vs secondary label question has also been discussed here, with answers from Stefan Kahl:
\nhttps://www.kaggle.com/c/birdclef-2021/discussion/231106
Short summary: If you treat rocpig and rocpig1 as one class (in the secondary labels, both representing the Rock Pigeon), then there are no secondary labels that are not also in the primary label set.
", "votes": null, "replies": [] }, { "id": 1309547, "author_name": "tonychenxyz", "author_url": "", "post_date": "05/16/2021 05:01:12", "content": "@jmreuter This is interesting. Thanks for the info!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1295050": "So we can ignore the sample of the classes of the metadata that are not in the train_soundscapes.", "1295144": "test set includes locations not present in train data hence restricting species may be a risky bet.", "1295170": "The metadata has 398 classes while the train_soundscapes only contains 49 classes. If the test set also has 49 classes then I may want to ignore the rest.", "1305565": "As a starter this is an interesting question to me. Is there an answer to it? Meaning, does the test data only contain classes from `train_soundscapes` or also some that are not present in `train_soundscapes`?", "1305575": "My bet is: train_soundscape < test_soundscape < metadata/snippets", "1305827": "I agree with @ilu000.\n\nSome species can be eliminated based on geographic range. Definitely no need to bother with Carolina Chickadee (carchi), for example; it is exceptionally unlikely to occur at any of the four test sites, as can be seen on the eBird range map:\nhttps://ebird.org/species/carchi\n\nOf the 398 species with short training audio, my estimates are that perhaps 315 of them are likely to occur at one or more of the four test sites at the time of year of the test soundscapes. But there's definitely slush in that estimate, as it is based on some assumptions that may or may not be accurate.", "1306180": "The train soundscapes don't cover all of the recording locations, but all species in the test set are present in the short train audio clips.", "1306322": "sohier Thanks. I made the same answer, but coming from you it may be trusted more ;)", "1306392": "Complete test set may have more sites. I wouldn't rule out any specific species. But there may for example be very rare ones with only single occurence or even zero occurence. \n\nEdit: seeing the comment below, seems I really need to read up a bit.", "1306583": "The complete test set has four sites; we've been given the approximate locations (lat/lon to two decimal points) as well as the recording dates (which I assume is a complete list?). Combining that information with data from eBird, there are some species that have a near-zero probability of being present or heard at any of the four sites. \n\nThe Carolina Chickadee is one that I have confidence as a birder will not be in any of the recordings for the four sites listed. Sapsucker Woods (New York) is the only test location remotely close to the Carolina Chickadee range, but Sapsucker Woods is firmly in Black-capped Chickadee territory. An unusually sharp boundary exists between the ranges for these two closely related species. Sapsucker Woods (being the home of the Cornell Lab of Ornithology) is ultra-heavily documented on eBird, yet a Carolina Chickadee has never been reported in Sapsucker Woods or its county, Tompkins County, New York. eBird does capture vagrants, and Tompkins County has some spectacular ones (Magnificent Frigatebird, for example), but literally no Carolina Chickadees.\n\nThe Carolina Chickadee is a case of a species that can be omitted with no harm in the context of the four test sites. I happen to think there are a number of others, though there’s a gray area of determining just how unlikely of a species to include. The fact that eBird coverage and effort isn’t equivalent for all four test regions should also be considered.", "1308108": "sohier Thanks for the answer. Does test set include species present only in primary labels in short train audio clips or those in secondary labels as well. Thanks!", "1308803": "From Stefan Kahl's [BirdCLEF2021: Exploring the data](https://www.kaggle.com/stefankahl/birdclef2021-exploring-the-data):\n\n> Please note, secondary lables only contain labels of species that are actually represented in the data set.\n\nI take that to mean that the short train audio clips do not list any secondary species that aren't already included in the primary list (though I technically haven't formally checked this). \n\nI also infer this to mean that other species might be in the background of the short train audio clips, but they wouldn't be listed. As Stefan also notes, \"lists of secondary labels are not very reliable\".", "1308822": "jmreuter I think some scondary label species are present in primary labels, while some don't.", "1309049": "tonychenxyz I just remembered that the primary vs secondary label question has also been discussed here, with answers from Stefan Kahl:\nhttps://www.kaggle.com/c/birdclef-2021/discussion/231106\n\nShort summary: If you treat rocpig and rocpig1 as one class (in the secondary labels, both representing the Rock Pigeon), then there are no secondary labels that are not also in the primary label set.", "1309547": "jmreuter This is interesting. Thanks for the info!" }, "source": "meta" }