{
  "id": 220379,
  "title": "post competition collaboration?",
  "url": "/competitions/rfcx-species-audio-detection/discussion/220379",
  "author_name": "",
  "post_date": "2021-02-18T05:54:15.875329300Z",
  "votes": 10,
  "comment_count": 8,
  "views": 0,
  "content": "<ul>\n<li><p>my solution is here:<br>\n<a href=\"https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\" target=\"_blank\">https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309</a></p></li>\n<li><p>0.937/0.939 (public/private) 5-fold</p></li>\n<li><p>note that I did not do any heavy augmentation. I did not use pesudo label.</p></li>\n</ul>\n<p>All kaggle audio competition will go back to \"data problem\", if you can create more and clean labels, you would basically get very good results. This is because data labeling is an issue in audio applications. Hence data creation is a basic skill you mask have in future audio competitions. The trick is to create data that has the same signal-to-noise characteristic in the test</p>\n<ul>\n<li>building upon my above solution, I need one to two helpers to:</li>\n</ul>\n<ol>\n<li>implement heavy augmentation.</li>\n<li>implement pesudo label.</li>\n<li>optional: transformer for self-supervised learning</li>\n<li>setup a GitHub repo (you need to oragnise the code, training log, etc, write simple document to repeat results, etc)</li>\n</ol>\n<p>what you would get</p>\n<ul>\n<li>I will supervise the implementation for 1 and 2. I estimate you should be able to get 0.970 in public/private LB (based on my initial experiments)</li>\n<li>there will be slack discussion group and video conference meeting</li>\n<li>you have early access to my code and results <br>\n(in simple words, I will discuss with you how to implement. You will do the implementations and show me the results and I will tell you how to improve, etc …)</li>\n</ul>\n<p>why am not doing it myself?</p>\n<ul>\n<li>I need to shift my attention to other competitions</li>\n</ul>\n<p>if you are interested please send me a private message via kaggle.<br>\nThanks!</p>",
  "messages": [
    {
      "id": "1208020",
      "postDate": "02/18/2021 05:54:15",
      "content": "<ul>\n<li><p>my solution is here:<br>\n<a href=\"https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\" target=\"_blank\">https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309</a></p></li>\n<li><p>0.937/0.939 (public/private) 5-fold</p></li>\n<li><p>note that I did not do any heavy augmentation. I did not use pesudo label.</p></li>\n</ul>\n<p>All kaggle audio competition will go back to \"data problem\", if you can create more and clean labels, you would basically get very good results. This is because data labeling is an issue in audio applications. Hence data creation is a basic skill you mask have in future audio competitions. The trick is to create data that has the same signal-to-noise characteristic in the test</p>\n<ul>\n<li>building upon my above solution, I need one to two helpers to:</li>\n</ul>\n<ol>\n<li>implement heavy augmentation.</li>\n<li>implement pesudo label.</li>\n<li>optional: transformer for self-supervised learning</li>\n<li>setup a GitHub repo (you need to oragnise the code, training log, etc, write simple document to repeat results, etc)</li>\n</ol>\n<p>what you would get</p>\n<ul>\n<li>I will supervise the implementation for 1 and 2. I estimate you should be able to get 0.970 in public/private LB (based on my initial experiments)</li>\n<li>there will be slack discussion group and video conference meeting</li>\n<li>you have early access to my code and results <br>\n(in simple words, I will discuss with you how to implement. You will do the implementations and show me the results and I will tell you how to improve, etc …)</li>\n</ul>\n<p>why am not doing it myself?</p>\n<ul>\n<li>I need to shift my attention to other competitions</li>\n</ul>\n<p>if you are interested please send me a private message via kaggle.<br>\nThanks!</p>",
      "rawMarkdown": "my solution is here:\nhttps://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\n\n- 0.937/0.939 (public/private) 5-fold\n- note that I did not do any heavy augmentation. I did not use pesudo label.\n\nAll kaggle audio competition will go back to \"data problem\", if you can create more and clean labels, you would basically get very good results. This is because data labeling is an issue in audio applications. Hence data creation is a basic skill you mask have in future audio competitions. The trick is to create data that has the same signal-to-noise characteristic in the test\n\n- building upon my above solution, I need one to two helpers to:\n1. implement heavy augmentation.\n2. implement pesudo label.\n3. optional: transformer for self-supervised learning\n4. setup a GitHub repo (you need to oragnise the code, training log, etc, write simple document to repeat results, etc)\n\nwhat you would get\n- I will supervise the implementation for 1 and 2. I estimate you should be able to get 0.970 in public/private LB (based on my initial experiments)\n- there will be slack discussion group and video conference meeting\n- you have early access to my code and results \n(in simple words, I will discuss with you how to implement. You will do the implementations and show me the results and I will tell you how to improve, etc ...)\n\nwhy am not doing it myself?\n- I need to shift my attention to other competitions\n\nif you are interested please send me a private message via kaggle.\nThanks!",
      "votes": null
    },
    {
      "id": "1209471",
      "postDate": "02/18/2021 23:14:42",
      "content": "<p>Thank you to those who have responded via kaggle. I have received a couple of emails.<br>\nPlease wait until this weekend, I will reply to all of you.</p>\n<p>I am now reading the top solutions and working out a plan on how the implementation of pseudo labeling (and self-supervised transformer) should be done. Please wait for a while.</p>\n<p>in particular, I am waiting for 2nd-rank solution. Hope Selim Seferbekov reveal something about LB 9.60 in 30 submission</p>",
      "rawMarkdown": "Thank you to those who have responded via kaggle. I have received a couple of emails.\nPlease wait until this weekend, I will reply to all of you.\n\nI am now reading the top solutions and working out a plan on how the implementation of pseudo labeling (and self-supervised transformer) should be done. Please wait for a while.\n\nin particular, I am waiting for 2nd-rank solution. Hope Selim Seferbekov reveal something about LB 9.60 in 30 submission",
      "votes": null
    },
    {
      "id": "1209618",
      "postDate": "02/19/2021 01:15:55",
      "content": "<p>OK, thank you. I have left a message to you and would like to do a lot of exploration/post-work to this competition!</p>",
      "rawMarkdown": "OK, thank you. I have left a message to you and would like to do a lot of exploration/post-work to this competition!",
      "votes": null
    },
    {
      "id": "1212763",
      "postDate": "02/21/2021 15:06:54",
      "content": "<p>Hi, it looks like I have not received your message yet, thanks~</p>",
      "rawMarkdown": "Hi, it looks like I have not received your message yet, thanks~",
      "votes": null
    },
    {
      "id": "1213178",
      "postDate": "02/21/2021 23:23:03",
      "content": "<p>I am in the same circumstance</p>",
      "rawMarkdown": "I am in the same circumstance",
      "votes": null
    },
    {
      "id": "1215467",
      "postDate": "02/23/2021 16:26:59",
      "content": "<p>i spend some time clean up the baseline code. It can now be found at the link:<br>\n<a href=\"https://drive.google.com/drive/folders/1jIKYnvAZvO2r8vqR8fAL5Nex16RlKHmw?usp=sharing\" target=\"_blank\">https://drive.google.com/drive/folders/1jIKYnvAZvO2r8vqR8fAL5Nex16RlKHmw?usp=sharing</a></p>\n<p>please see folder 2020-02-23. there is a read me file for instruction.</p>\n<pre><code>1. this code implements 5-fold ensemble of resnet34-conv-classifier (stride-16) that achieve:\n   ensemble-fold01234 = 0.9452/0.9426 (private/public LB)\n\nsee https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\n</code></pre>",
      "rawMarkdown": "i spend some time clean up the baseline code. It can now be found at the link:\nhttps://drive.google.com/drive/folders/1jIKYnvAZvO2r8vqR8fAL5Nex16RlKHmw?usp=sharing\n\nplease see folder 2020-02-23. there is a read me file for instruction.\n```\n1. this code implements 5-fold ensemble of resnet34-conv-classifier (stride-16) that achieve:\n   ensemble-fold01234 = 0.9452/0.9426 (private/public LB)\n\nsee https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\n```",
      "votes": null
    },
    {
      "id": "1215606",
      "postDate": "02/23/2021 19:23:16",
      "content": "<p>I added a new delivery, please see folder \"2020-02-24\". It trains another model for mobilenetv3.<br>\nwith the ensemble of mobilenetv3+resnet34, it achieves  Lb 0.9501/0.9478 (private/public) which is actually a gold model. This completes the baseline solution for the pseudo-label experiment.</p>\n<p><img src=\"https://i.ibb.co/Qms777V/Selection-166.png\" alt=\"\"></p>",
      "rawMarkdown": "I added a new delivery, please see folder \"2020-02-24\". It trains another model for mobilenetv3.\nwith the ensemble of mobilenetv3+resnet34, it achieves  Lb 0.9501/0.9478 (private/public) which is actually a gold model. This completes the baseline solution for the pseudo-label experiment.\n\n![](https://i.ibb.co/Qms777V/Selection-166.png)",
      "votes": null
    },
    {
      "id": "1216327",
      "postDate": "02/24/2021 08:00:59",
      "content": "<p>update: \"folder 2020-20-24a\" for plugin starter code for pesudo label <br>\n<img src=\"https://i.ibb.co/6y4WtNj/Selection-176.png\" alt=\"\"><br>\n<img src=\"https://i.ibb.co/FJH2q5d/Selection-175.png\" alt=\"\"></p>",
      "rawMarkdown": "update: \"folder 2020-20-24a\" for plugin starter code for pesudo label \n![](https://i.ibb.co/6y4WtNj/Selection-176.png)\n![](https://i.ibb.co/FJH2q5d/Selection-175.png)",
      "votes": null
    },
    {
      "id": "1216329",
      "postDate": "02/24/2021 08:01:55",
      "content": "<p>sorry for being late: you can join the slack group from this link as well:<br>\n<a href=\"https://join.slack.com/t/newworkspace-s6n3109/shared_invite/zt-mpv3wmti-iKZX8TnN2RsuHcoqR~bOsg\" target=\"_blank\">https://join.slack.com/t/newworkspace-s6n3109/shared_invite/zt-mpv3wmti-iKZX8TnN2RsuHcoqR~bOsg</a></p>",
      "rawMarkdown": "sorry for being late: you can join the slack group from this link as well:\nhttps://join.slack.com/t/newworkspace-s6n3109/shared_invite/zt-mpv3wmti-iKZX8TnN2RsuHcoqR~bOsg",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1209471,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "02/18/2021 23:14:42",
      "content": "<p>Thank you to those who have responded via kaggle. I have received a couple of emails.<br>\nPlease wait until this weekend, I will reply to all of you.</p>\n<p>I am now reading the top solutions and working out a plan on how the implementation of pseudo labeling (and self-supervised transformer) should be done. Please wait for a while.</p>\n<p>in particular, I am waiting for 2nd-rank solution. Hope Selim Seferbekov reveal something about LB 9.60 in 30 submission</p>",
      "votes": null,
      "replies": [
        {
          "id": 1212763,
          "author_name": "wubinbai",
          "author_url": "",
          "post_date": "02/21/2021 15:06:54",
          "content": "<p>Hi, it looks like I have not received your message yet, thanks~</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1213178,
          "author_name": "felipebihaiek",
          "author_url": "",
          "post_date": "02/21/2021 23:23:03",
          "content": "<p>I am in the same circumstance</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1216329,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "02/24/2021 08:01:55",
          "content": "<p>sorry for being late: you can join the slack group from this link as well:<br>\n<a href=\"https://join.slack.com/t/newworkspace-s6n3109/shared_invite/zt-mpv3wmti-iKZX8TnN2RsuHcoqR~bOsg\" target=\"_blank\">https://join.slack.com/t/newworkspace-s6n3109/shared_invite/zt-mpv3wmti-iKZX8TnN2RsuHcoqR~bOsg</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1209618,
      "author_name": "wubinbai",
      "author_url": "",
      "post_date": "02/19/2021 01:15:55",
      "content": "<p>OK, thank you. I have left a message to you and would like to do a lot of exploration/post-work to this competition!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1215467,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "02/23/2021 16:26:59",
      "content": "<p>i spend some time clean up the baseline code. It can now be found at the link:<br>\n<a href=\"https://drive.google.com/drive/folders/1jIKYnvAZvO2r8vqR8fAL5Nex16RlKHmw?usp=sharing\" target=\"_blank\">https://drive.google.com/drive/folders/1jIKYnvAZvO2r8vqR8fAL5Nex16RlKHmw?usp=sharing</a></p>\n<p>please see folder 2020-02-23. there is a read me file for instruction.</p>\n<pre><code>1. this code implements 5-fold ensemble of resnet34-conv-classifier (stride-16) that achieve:\n   ensemble-fold01234 = 0.9452/0.9426 (private/public LB)\n\nsee https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\n</code></pre>",
      "votes": null,
      "replies": [
        {
          "id": 1215606,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "02/23/2021 19:23:16",
          "content": "<p>I added a new delivery, please see folder \"2020-02-24\". It trains another model for mobilenetv3.<br>\nwith the ensemble of mobilenetv3+resnet34, it achieves  Lb 0.9501/0.9478 (private/public) which is actually a gold model. This completes the baseline solution for the pseudo-label experiment.</p>\n<p><img src=\"https://i.ibb.co/Qms777V/Selection-166.png\" alt=\"\"></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1216327,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "02/24/2021 08:00:59",
          "content": "<p>update: \"folder 2020-20-24a\" for plugin starter code for pesudo label <br>\n<img src=\"https://i.ibb.co/6y4WtNj/Selection-176.png\" alt=\"\"><br>\n<img src=\"https://i.ibb.co/FJH2q5d/Selection-175.png\" alt=\"\"></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1208020": "my solution is here:\nhttps://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\n\n- 0.937/0.939 (public/private) 5-fold\n- note that I did not do any heavy augmentation. I did not use pesudo label.\n\nAll kaggle audio competition will go back to \"data problem\", if you can create more and clean labels, you would basically get very good results. This is because data labeling is an issue in audio applications. Hence data creation is a basic skill you mask have in future audio competitions. The trick is to create data that has the same signal-to-noise characteristic in the test\n\n- building upon my above solution, I need one to two helpers to:\n1. implement heavy augmentation.\n2. implement pesudo label.\n3. optional: transformer for self-supervised learning\n4. setup a GitHub repo (you need to oragnise the code, training log, etc, write simple document to repeat results, etc)\n\nwhat you would get\n- I will supervise the implementation for 1 and 2. I estimate you should be able to get 0.970 in public/private LB (based on my initial experiments)\n- there will be slack discussion group and video conference meeting\n- you have early access to my code and results \n(in simple words, I will discuss with you how to implement. You will do the implementations and show me the results and I will tell you how to improve, etc ...)\n\nwhy am not doing it myself?\n- I need to shift my attention to other competitions\n\nif you are interested please send me a private message via kaggle.\nThanks!",
    "1209471": "Thank you to those who have responded via kaggle. I have received a couple of emails.\nPlease wait until this weekend, I will reply to all of you.\n\nI am now reading the top solutions and working out a plan on how the implementation of pseudo labeling (and self-supervised transformer) should be done. Please wait for a while.\n\nin particular, I am waiting for 2nd-rank solution. Hope Selim Seferbekov reveal something about LB 9.60 in 30 submission",
    "1209618": "OK, thank you. I have left a message to you and would like to do a lot of exploration/post-work to this competition!",
    "1212763": "Hi, it looks like I have not received your message yet, thanks~",
    "1213178": "I am in the same circumstance",
    "1215467": "i spend some time clean up the baseline code. It can now be found at the link:\nhttps://drive.google.com/drive/folders/1jIKYnvAZvO2r8vqR8fAL5Nex16RlKHmw?usp=sharing\n\nplease see folder 2020-02-23. there is a read me file for instruction.\n```\n1. this code implements 5-fold ensemble of resnet34-conv-classifier (stride-16) that achieve:\n   ensemble-fold01234 = 0.9452/0.9426 (private/public LB)\n\nsee https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220309\n```",
    "1215606": "I added a new delivery, please see folder \"2020-02-24\". It trains another model for mobilenetv3.\nwith the ensemble of mobilenetv3+resnet34, it achieves  Lb 0.9501/0.9478 (private/public) which is actually a gold model. This completes the baseline solution for the pseudo-label experiment.\n\n![](https://i.ibb.co/Qms777V/Selection-166.png)",
    "1216327": "update: \"folder 2020-20-24a\" for plugin starter code for pesudo label \n![](https://i.ibb.co/6y4WtNj/Selection-176.png)\n![](https://i.ibb.co/FJH2q5d/Selection-175.png)",
    "1216329": "sorry for being late: you can join the slack group from this link as well:\nhttps://join.slack.com/t/newworkspace-s6n3109/shared_invite/zt-mpv3wmti-iKZX8TnN2RsuHcoqR~bOsg"
  },
  "source": "meta"
}