{ "id": 243312, "title": "42 solution with code", "url": "/competitions/birdclef-2021/writeups/mocking-bird-42-solution-with-code", "author_name": "", "post_date": "2021-06-02T02:31:52.980Z", "votes": 9, "comment_count": 2, "views": 0, "content": "

thanks to the hosts, kaggle
\nthanks to @kkiller and his awosome code

\n

https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-training-colab

\n

3 step

\n
    \n
  1. train different model (resnet34, resnet50,resnest50,resnext50_32x4d)
    2. \n
  2. find best ensemble model, depend on train_soundscapes
    3. \n
  3. post preprocess, use season and position to filter bird
    4. \n
\n

train

\n

I use kkiller public datasets, and I try to add some augument, but not increase much, less 0.01
\nI add second label, with prob 0.3, have some improvement, I study a lot last year's solution

\n

https://www.kaggle.com/c/birdsong-recognition/discussion/183208
\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183269
\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183199
\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183204

\n

I also tried effientnet b0~b5, not work, I don't known why, wished players can give some advice

\n

ensemble

\n

I trained diffent model, and use train_soundscapes f1 to check best epoch
\nI choose 2 sulotion, one for best local f1, and one add some random, use model not sort by local f1

\n

post preprocess

\n

I use uber h3 calculate Latitude and Longitude to get segment id, and use train_metadata.csv to get bird belong which segment, this give 0.01 improve

\n

finally

\n

thanks all, I am NLPer, not very familier with sound engineer, but kindly public code helped me a lot

\n

my inference code:

\n

https://www.kaggle.com/jt120lz/clean-fast-simple-bird-identifier-inference

", "messages": [ { "id": "1332262", "postDate": "06/02/2021 02:28:19", "content": "

thanks to the hosts, kaggle
\nthanks to @kkiller and his awosome code

\n

https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-training-colab

\n

3 step

\n
    \n
  1. train different model (resnet34, resnet50,resnest50,resnext50_32x4d)
    2. \n
  2. find best ensemble model, depend on train_soundscapes
    3. \n
  3. post preprocess, use season and position to filter bird
    4. \n
\n

train

\n

I use kkiller public datasets, and I try to add some augument, but not increase much, less 0.01
\nI add second label, with prob 0.3, have some improvement, I study a lot last year's solution

\n

https://www.kaggle.com/c/birdsong-recognition/discussion/183208
\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183269
\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183199
\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183204

\n

I also tried effientnet b0~b5, not work, I don't known why, wished players can give some advice

\n

ensemble

\n

I trained diffent model, and use train_soundscapes f1 to check best epoch
\nI choose 2 sulotion, one for best local f1, and one add some random, use model not sort by local f1

\n

post preprocess

\n

I use uber h3 calculate Latitude and Longitude to get segment id, and use train_metadata.csv to get bird belong which segment, this give 0.01 improve

\n

finally

\n

thanks all, I am NLPer, not very familier with sound engineer, but kindly public code helped me a lot

\n

my inference code:

\n

https://www.kaggle.com/jt120lz/clean-fast-simple-bird-identifier-inference

", "rawMarkdown": "thanks to the hosts, kaggle\nthanks to @kkiller and his awosome code\n\nhttps://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-training-colab\n\n## 3 step\n\n1. train different model (resnet34, resnet50,resnest50,resnext50_32x4d)\n2. find best ensemble model, depend on train_soundscapes\n3. post preprocess, use season and position to filter bird\n\n\n### train\n\nI use kkiller public datasets, and I try to add some augument, but not increase much, less 0.01\nI add second label, with prob 0.3, have some improvement, I study a lot last year's solution\n\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183208\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183269\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183199\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183204\n\nI also tried effientnet b0~b5, not work, I don't known why, wished players can give some advice\n\n### ensemble\n\nI trained diffent model, and use train_soundscapes f1 to check best epoch\nI choose 2 sulotion, one for best local f1, and one add some random, use model not sort by local f1\n\n### post preprocess\n\nI use uber h3 calculate Latitude and Longitude to get segment id, and use train_metadata.csv to get bird belong which segment, this give 0.01 improve\n\n### finally\n\n thanks all, I am NLPer, not very familier with sound engineer, but kindly public code helped me a lot\n\nmy inference code:\n\nhttps://www.kaggle.com/jt120lz/clean-fast-simple-bird-identifier-inference", "votes": null }, { "id": "1332275", "postDate": "06/02/2021 02:48:39", "content": "

Congrat and thanks for the quick write-up!
\nHow did u select the best ensemble model and what ensemble techniques did u try?

", "rawMarkdown": "Congrat and thanks for the quick write-up!\nHow did u select the best ensemble model and what ensemble techniques did u try?", "votes": null }, { "id": "1332361", "postDate": "06/02/2021 04:56:53", "content": "

just for loop, choose best f1 on train_soundscapes

", "rawMarkdown": "just for loop, choose best f1 on train_soundscapes", "votes": null } ], "comments": [ { "id": 1332275, "author_name": "alexlwh", "author_url": "", "post_date": "06/02/2021 02:48:39", "content": "

Congrat and thanks for the quick write-up!
\nHow did u select the best ensemble model and what ensemble techniques did u try?

", "votes": null, "replies": [ { "id": 1332361, "author_name": "jt120lz", "author_url": "", "post_date": "06/02/2021 04:56:53", "content": "

just for loop, choose best f1 on train_soundscapes

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1332262": "thanks to the hosts, kaggle\nthanks to @kkiller and his awosome code\n\nhttps://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-training-colab\n\n## 3 step\n\n1. train different model (resnet34, resnet50,resnest50,resnext50_32x4d)\n2. find best ensemble model, depend on train_soundscapes\n3. post preprocess, use season and position to filter bird\n\n\n### train\n\nI use kkiller public datasets, and I try to add some augument, but not increase much, less 0.01\nI add second label, with prob 0.3, have some improvement, I study a lot last year's solution\n\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183208\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183269\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183199\nhttps://www.kaggle.com/c/birdsong-recognition/discussion/183204\n\nI also tried effientnet b0~b5, not work, I don't known why, wished players can give some advice\n\n### ensemble\n\nI trained diffent model, and use train_soundscapes f1 to check best epoch\nI choose 2 sulotion, one for best local f1, and one add some random, use model not sort by local f1\n\n### post preprocess\n\nI use uber h3 calculate Latitude and Longitude to get segment id, and use train_metadata.csv to get bird belong which segment, this give 0.01 improve\n\n### finally\n\n thanks all, I am NLPer, not very familier with sound engineer, but kindly public code helped me a lot\n\nmy inference code:\n\nhttps://www.kaggle.com/jt120lz/clean-fast-simple-bird-identifier-inference", "1332275": "Congrat and thanks for the quick write-up!\nHow did u select the best ensemble model and what ensemble techniques did u try?", "1332361": "just for loop, choose best f1 on train_soundscapes" }, "source": "meta" }