{ "id": 230539, "title": "Starter and some thoughts", "url": "/competitions/birdclef-2021/discussion/230539", "author_name": "Hidehisa Arai", "post_date": "2021-04-04T10:54:34.599000", "votes": 52, "comment_count": 5, "views": 0, "content": "

Hi all,

\n

I've shared a simple baseline Notebook here
\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-inference-birdclef2021-starter

\n

I'll also make training Notebook soon.

\n

As one of the participants of the last Birdcall competition, I'll put some thoughts on this competition here.

\n
    \n
  1. The setting are similar.
    2. \n
\n

We need to create an audio tagging system for soundscapes with web-crawled audio clips. This is the same setting as that of the last Birdcall's competition. The performance is measured with chunk level sample f1, however, we don't have chunk level annotation for the training data(train_short_audio). This is also the same.

\n
    \n
  1. Big difference - we have some soundscapes this time.
    2. \n
\n

We can use train_soundscapes this time. For example, we can use these files for validation, or we can cut some part out of these clips as background noise.

\n
    \n
  1. We have location information
    2. \n
\n

This time we can use the location information. We can use these information to make our models more focused on the specific sites.

", "messages": [ { "id": 1262477, "postDate": "2021-04-04T10:54:34.600Z", "content": "

Hi all,

\n

I've shared a simple baseline Notebook here
\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-inference-birdclef2021-starter

\n

I'll also make training Notebook soon.

\n

As one of the participants of the last Birdcall competition, I'll put some thoughts on this competition here.

\n
    \n
  1. The setting are similar.
    2. \n
\n

We need to create an audio tagging system for soundscapes with web-crawled audio clips. This is the same setting as that of the last Birdcall's competition. The performance is measured with chunk level sample f1, however, we don't have chunk level annotation for the training data(train_short_audio). This is also the same.

\n
    \n
  1. Big difference - we have some soundscapes this time.
    2. \n
\n

We can use train_soundscapes this time. For example, we can use these files for validation, or we can cut some part out of these clips as background noise.

\n
    \n
  1. We have location information
    2. \n
\n

This time we can use the location information. We can use these information to make our models more focused on the specific sites.

", "rawMarkdown": "Hi all,\n\nI've shared a simple baseline Notebook here\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-inference-birdclef2021-starter\n\nI'll also make training Notebook soon.\n\nAs one of the participants of the last Birdcall competition, I'll put some thoughts on this competition here.\n\n1. The setting are similar.\n\nWe need to create an audio tagging system for soundscapes with web-crawled audio clips. This is the same setting as that of the last Birdcall's competition. The performance is measured with chunk level sample f1, however, we don't have chunk level annotation for the training data(`train_short_audio`). This is also the same.\n\n2. Big difference - we have some soundscapes this time.\n\nWe can use `train_soundscapes` this time. For example, we can use these files for validation, or we can cut some part out of these clips as background noise.\n\n3. We have location information\n\nThis time we can use the location information. We can use these information to make our models more focused on the specific sites.\n", "votes": 51 }, { "id": 1264091, "postDate": "2021-04-05T21:40:27.647Z", "content": "

Here's the training notebook
\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-training-birdclef2021-starter

", "rawMarkdown": "Here's the training notebook\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-training-birdclef2021-starter", "votes": 5 }, { "id": 1262503, "postDate": "2021-04-04T11:42:10.060Z", "content": "

Good point about location, this is new.

", "rawMarkdown": "Good point about location, this is new.", "votes": 4, "replies": [ { "id": 1262508, "postDate": "2021-04-04T11:48:42.777Z", "content": "

I think this will be one of the key in this competition.
\nI'll work on checking the difference between sites, and hopefully publish some EDA notebook(s) on it

", "rawMarkdown": "I think this will be one of the key in this competition.\nI'll work on checking the difference between sites, and hopefully publish some EDA notebook(s) on it", "votes": 5 } ] }, { "id": 1329238, "postDate": "2021-05-31T01:53:51.243Z", "content": "

Thanks for the starter Hidehisa. This is only my third Kaggle comp and some of your methods are a bit over my head, but it's been a great learning exercise. I'm hoping to combine it with some temporal and spatial models.

", "rawMarkdown": "Thanks for the starter Hidehisa. This is only my third Kaggle comp and some of your methods are a bit over my head, but it's been a great learning exercise. I'm hoping to combine it with some temporal and spatial models." }, { "id": 1262576, "postDate": "2021-04-04T13:13:25.920Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1264091, "author_name": "Hidehisa Arai", "author_url": "", "post_date": "2021-04-05T21:40:27.647000", "content": "

Here's the training notebook
\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-training-birdclef2021-starter

", "votes": 5, "replies": [] }, { "id": 1262503, "author_name": "CPMP", "author_url": "", "post_date": "2021-04-04T11:42:10.060000", "content": "

Good point about location, this is new.

", "votes": 4, "replies": [ { "id": 1262508, "author_name": "Hidehisa Arai", "author_url": "", "post_date": "2021-04-04T11:48:42.777000", "content": "

I think this will be one of the key in this competition.
\nI'll work on checking the difference between sites, and hopefully publish some EDA notebook(s) on it

", "votes": 5, "replies": [] } ] }, { "id": 1329238, "author_name": "Olly Powell", "author_url": "", "post_date": "2021-05-31T01:53:51.243000", "content": "

Thanks for the starter Hidehisa. This is only my third Kaggle comp and some of your methods are a bit over my head, but it's been a great learning exercise. I'm hoping to combine it with some temporal and spatial models.

", "votes": 0, "replies": [] }, { "id": 1262576, "author_name": "", "author_url": "", "post_date": "2021-04-04T13:13:25.920000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1262477": "Hi all,\n\nI've shared a simple baseline Notebook here\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-inference-birdclef2021-starter\n\nI'll also make training Notebook soon.\n\nAs one of the participants of the last Birdcall competition, I'll put some thoughts on this competition here.\n\n1. The setting are similar.\n\nWe need to create an audio tagging system for soundscapes with web-crawled audio clips. This is the same setting as that of the last Birdcall's competition. The performance is measured with chunk level sample f1, however, we don't have chunk level annotation for the training data(`train_short_audio`). This is also the same.\n\n2. Big difference - we have some soundscapes this time.\n\nWe can use `train_soundscapes` this time. For example, we can use these files for validation, or we can cut some part out of these clips as background noise.\n\n3. We have location information\n\nThis time we can use the location information. We can use these information to make our models more focused on the specific sites.\n", "1264091": "Here's the training notebook\nhttps://www.kaggle.com/hidehisaarai1213/pytorch-training-birdclef2021-starter", "1262503": "Good point about location, this is new.", "1329238": "Thanks for the starter Hidehisa. This is only my third Kaggle comp and some of your methods are a bit over my head, but it's been a great learning exercise. I'm hoping to combine it with some temporal and spatial models.", "1262576": "" } }