{ "id": 327575, "title": "70th place solution.", "url": "/competitions/birdclef-2022/discussion/327575", "author_name": "yoyobar", "post_date": "2022-05-27T23:36:15.569000", "votes": 4, "comment_count": 0, "views": 0, "content": "

This was an good competition for me and I have learned a lot from it. thank the hosts and my teammates, and the kagglers who shared ideas!

\n

BirdCLEF 2022 submit 70th pubilc:0.77,private:0.71

\n

there is our used technique

\n

useful:

\n
    \n
  1. Cut the input into 4 pieces and predict and solve for the mean.
    2. \n
  2. used 5 second to training.
    3. \n
  3. Mix two audio to one.
    4. \n
  4. Gaussian noise and pink noise.
    5. \n
  5. using a percentage-based threshold method (it helped we to do not test LB so many time. [1]
    6. \n
\n

used it but don't know if it works:

\n
    \n
  1. PCEN.
    2. \n
  2. AdamW.
    3. \n
  3. sample equalization.
    4. \n
\n

useless :

\n
    \n
  1. focal loss (Both CV and LB have dropped)
    2. \n
\n

In our final submission ,it got CV:0.82 ,public score :0.77 private:0.7124.
\nAll training is done on my 3080ti.

\n
#
\n

non-technical things

\n

This is my the first time competition in my life ,and I got some friend form that.

\n

I have met obstacle for example such as many submissions got 49~51 public score [2] and I did not know why, I joined someone else's team,Because of some accidents [3],
\nBut it all worked out in the end.

\n

[1] 2nd place solution.
\n[2] The lower threshhold has good score.
\n[3] Communication within the team.

", "messages": [ { "id": 1803540, "postDate": "2022-05-27T23:36:15.570Z", "content": "

This was an good competition for me and I have learned a lot from it. thank the hosts and my teammates, and the kagglers who shared ideas!

\n

BirdCLEF 2022 submit 70th pubilc:0.77,private:0.71

\n

there is our used technique

\n

useful:

\n
    \n
  1. Cut the input into 4 pieces and predict and solve for the mean.
    2. \n
  2. used 5 second to training.
    3. \n
  3. Mix two audio to one.
    4. \n
  4. Gaussian noise and pink noise.
    5. \n
  5. using a percentage-based threshold method (it helped we to do not test LB so many time. [1]
    6. \n
\n

used it but don't know if it works:

\n
    \n
  1. PCEN.
    2. \n
  2. AdamW.
    3. \n
  3. sample equalization.
    4. \n
\n

useless :

\n
    \n
  1. focal loss (Both CV and LB have dropped)
    2. \n
\n

In our final submission ,it got CV:0.82 ,public score :0.77 private:0.7124.
\nAll training is done on my 3080ti.

\n
#
\n

non-technical things

\n

This is my the first time competition in my life ,and I got some friend form that.

\n

I have met obstacle for example such as many submissions got 49~51 public score [2] and I did not know why, I joined someone else's team,Because of some accidents [3],
\nBut it all worked out in the end.

\n

[1] 2nd place solution.
\n[2] The lower threshhold has good score.
\n[3] Communication within the team.

", "rawMarkdown": "This was an good competition for me and I have learned a lot from it. thank the hosts and my teammates, and the kagglers who shared ideas!\n\n[BirdCLEF 2022 submit 70th pubilc:0.77,private:0.71](https://www.kaggle.com/code/yoyobar/birdclef-2022-submit-70th-pubilc-0-77-private-0-71/settings?scriptVersionId=95983080)\n\nthere is our used technique\n\nuseful:\n1. Cut the input into 4 pieces and predict and solve for the mean.\n2. used 5 second to training.\n3. Mix two audio to one.\n4. Gaussian noise and pink noise.\n5. using a percentage-based threshold method (it helped we to do not test LB so many time. [1]\n\nused it but don't know if it works:\n1. PCEN.\n2. AdamW.\n3. sample equalization.\n\nuseless :\n1. focal loss (Both CV and LB have dropped)\n\nIn our final submission ,it got CV:0.82 ,public score :0.77 private:0.7124.\nAll training is done on my 3080ti.\n########################################################\nnon-technical things\n\nThis is my the first time competition in my life ,and I got some friend form that.\n\nI have met obstacle for example such as many submissions got 49~51 public score [2] and I did not know why, I joined someone else's team,Because of some accidents [3], \nBut it all worked out in the end.\n\n[1] [2nd place solution.](https://www.kaggle.com/competitions/birdclef-2021/discussion/243463)\n[2] [The lower threshhold has good score.](https://www.kaggle.com/competitions/birdclef-2022/discussion/318999)\n[3] [Communication within the team.](https://www.kaggle.com/competitions/birdclef-2022/discussion/321584)\n", "votes": 4 } ], "comments": [], "raw_markdown_by_id": { "1803540": "This was an good competition for me and I have learned a lot from it. thank the hosts and my teammates, and the kagglers who shared ideas!\n\n[BirdCLEF 2022 submit 70th pubilc:0.77,private:0.71](https://www.kaggle.com/code/yoyobar/birdclef-2022-submit-70th-pubilc-0-77-private-0-71/settings?scriptVersionId=95983080)\n\nthere is our used technique\n\nuseful:\n1. Cut the input into 4 pieces and predict and solve for the mean.\n2. used 5 second to training.\n3. Mix two audio to one.\n4. Gaussian noise and pink noise.\n5. using a percentage-based threshold method (it helped we to do not test LB so many time. [1]\n\nused it but don't know if it works:\n1. PCEN.\n2. AdamW.\n3. sample equalization.\n\nuseless :\n1. focal loss (Both CV and LB have dropped)\n\nIn our final submission ,it got CV:0.82 ,public score :0.77 private:0.7124.\nAll training is done on my 3080ti.\n########################################################\nnon-technical things\n\nThis is my the first time competition in my life ,and I got some friend form that.\n\nI have met obstacle for example such as many submissions got 49~51 public score [2] and I did not know why, I joined someone else's team,Because of some accidents [3], \nBut it all worked out in the end.\n\n[1] [2nd place solution.](https://www.kaggle.com/competitions/birdclef-2021/discussion/243463)\n[2] [The lower threshhold has good score.](https://www.kaggle.com/competitions/birdclef-2022/discussion/318999)\n[3] [Communication within the team.](https://www.kaggle.com/competitions/birdclef-2022/discussion/321584)\n" } }