As it turned out, resampling the train set to 48 kHz and concatenating the files by folder turned out to not be such a great idea 😄
\n\nWe now know though, thanks to the kindness of @stefankahl, that the test set should have a sample rate of 32 kHz across the board. That means that training on data with a higher sampling rate probably does not make sense.
\n\nAnyhow - here is the entire train set resampled at 32 kHz with the file and directory structure preserved. Next stop - building datasets that will split data into train and validation set by recordings stratified by ebird_code.
", "messages": [ { "id": "894800", "postDate": "06/20/2020 19:05:48", "content": "As it turned out, resampling the train set to 48 kHz and concatenating the files by folder turned out to not be such a great idea 😄
\n\nWe now know though, thanks to the kindness of @stefankahl, that the test set should have a sample rate of 32 kHz across the board. That means that training on data with a higher sampling rate probably does not make sense.
\n\nAnyhow - here is the entire train set resampled at 32 kHz with the file and directory structure preserved. Next stop - building datasets that will split data into train and validation set by recordings stratified by ebird_code.
", "rawMarkdown": "As it turned out, resampling the train set to 48 kHz and concatenating the files by folder turned out [to not be such a great idea](https://www.kaggle.com/c/birdsong-recognition/discussion/159943) 😄\n\nWe now know though, thanks to the kindness of @stefankahl, that the test set should have a sample rate of 32 kHz across the board. That means that training on data with a higher sampling rate probably does not make sense.\n\nAnyhow - [here](https://storage.googleapis.com/birdcall_competition/train_resampled.zip) is the entire train set resampled at 32 kHz with the file and directory structure preserved. Next stop - building datasets that will split data into train and validation set by recordings stratified by ebird_code.", "votes": null }, { "id": "900034", "postDate": "06/24/2020 15:28:59", "content": "Good one!
", "rawMarkdown": "Good one!", "votes": null }, { "id": "966876", "postDate": "08/11/2020 17:58:01", "content": "Thanks for your effort @radek1 By the way data is large so double check that you are OK with GCS network egress prices.
", "rawMarkdown": "Thanks for your effort @radek1 By the way data is large so double check that you are OK with [GCS network egress prices](https://cloud.google.com/storage/pricing#network-egress).", "votes": null } ], "comments": [ { "id": 966876, "author_name": "vecxoz", "author_url": "", "post_date": "08/11/2020 17:58:01", "content": "Thanks for your effort @radek1 By the way data is large so double check that you are OK with GCS network egress prices.
", "votes": null, "replies": [] }, { "id": 900034, "author_name": "mdselimreza", "author_url": "", "post_date": "06/24/2020 15:28:59", "content": "Good one!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "894800": "As it turned out, resampling the train set to 48 kHz and concatenating the files by folder turned out [to not be such a great idea](https://www.kaggle.com/c/birdsong-recognition/discussion/159943) 😄\n\nWe now know though, thanks to the kindness of @stefankahl, that the test set should have a sample rate of 32 kHz across the board. That means that training on data with a higher sampling rate probably does not make sense.\n\nAnyhow - [here](https://storage.googleapis.com/birdcall_competition/train_resampled.zip) is the entire train set resampled at 32 kHz with the file and directory structure preserved. Next stop - building datasets that will split data into train and validation set by recordings stratified by ebird_code.", "900034": "Good one!", "966876": "Thanks for your effort @radek1 By the way data is large so double check that you are OK with [GCS network egress prices](https://cloud.google.com/storage/pricing#network-egress)." }, "source": "meta" }