{
  "id": 307045,
  "title": "Some ideas that worked/not worked for me",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/307045",
  "author_name": "Vlad Vaduva",
  "post_date": "2022-02-12T08:44:03.389000",
  "votes": 60,
  "comment_count": 33,
  "views": 0,
  "content": "<p><strong>Ideas that worked:</strong></p>\n<ul>\n<li>Using small architecture like Yolo5s6</li>\n<li>Adam instead of SGD</li>\n<li>Larger training image size than the default images sizes</li>\n<li>Adjusting augmentation (Here is important to adapt to the model size, augmentations for Yolo5s6 are not enough for Yolo5l6 in order to generalize) </li>\n<li>Retraining with whole data after selecting the proper model</li>\n<li>Adjusting learning rate to correspond with the optimizer characteristics and batch size (wandb is very useful tool for observing the model evolution on training/valid)</li>\n<li>Tracking (adjusting parameters may help more)</li>\n</ul>\n<p><strong>Ideas that did not work:</strong></p>\n<ul>\n<li>Ensembles (so far). I am going to spend the remaining submissions optimizing ensembles techniques</li>\n<li>Adding no-label images to training data in a proportion of 5-10% of the total training data</li>\n<li>CLAHE</li>\n<li>GAN techniques for increasing COTS in training data</li>\n<li>SAHI  techniques did not worked for me, but I love the idea. Probably with more time invested in this, good results can be achieved</li>\n<li>YOLOX got me lower performance comparative with YOLO5</li>\n</ul>\n<p>Good luck to everybody and congratulations to all participants for coming up with a tone of interesting ideas! This is one of the most interesting competitions from the last year </p>",
  "messages": [
    {
      "id": 1686657,
      "postDate": "2022-02-12T08:44:03.390Z",
      "content": "<p><strong>Ideas that worked:</strong></p>\n<ul>\n<li>Using small architecture like Yolo5s6</li>\n<li>Adam instead of SGD</li>\n<li>Larger training image size than the default images sizes</li>\n<li>Adjusting augmentation (Here is important to adapt to the model size, augmentations for Yolo5s6 are not enough for Yolo5l6 in order to generalize) </li>\n<li>Retraining with whole data after selecting the proper model</li>\n<li>Adjusting learning rate to correspond with the optimizer characteristics and batch size (wandb is very useful tool for observing the model evolution on training/valid)</li>\n<li>Tracking (adjusting parameters may help more)</li>\n</ul>\n<p><strong>Ideas that did not work:</strong></p>\n<ul>\n<li>Ensembles (so far). I am going to spend the remaining submissions optimizing ensembles techniques</li>\n<li>Adding no-label images to training data in a proportion of 5-10% of the total training data</li>\n<li>CLAHE</li>\n<li>GAN techniques for increasing COTS in training data</li>\n<li>SAHI  techniques did not worked for me, but I love the idea. Probably with more time invested in this, good results can be achieved</li>\n<li>YOLOX got me lower performance comparative with YOLO5</li>\n</ul>\n<p>Good luck to everybody and congratulations to all participants for coming up with a tone of interesting ideas! This is one of the most interesting competitions from the last year </p>",
      "rawMarkdown": "**Ideas that worked:**\n- Using small architecture like Yolo5s6\n- Adam instead of SGD\n- Larger training image size than the default images sizes\n- Adjusting augmentation (Here is important to adapt to the model size, augmentations for Yolo5s6 are not enough for Yolo5l6 in order to generalize) \n- Retraining with whole data after selecting the proper model\n- Adjusting learning rate to correspond with the optimizer characteristics and batch size (wandb is very useful tool for observing the model evolution on training/valid)\n- Tracking (adjusting parameters may help more)\n\n**Ideas that did not work:**\n- Ensembles (so far). I am going to spend the remaining submissions optimizing ensembles techniques\n- Adding no-label images to training data in a proportion of 5-10% of the total training data\n- CLAHE\n- GAN techniques for increasing COTS in training data\n- SAHI  techniques did not worked for me, but I love the idea. Probably with more time invested in this, good results can be achieved\n- YOLOX got me lower performance comparative with YOLO5\n\nGood luck to everybody and congratulations to all participants for coming up with a tone of interesting ideas! This is one of the most interesting competitions from the last year ",
      "votes": 60
    },
    {
      "id": 1686942,
      "postDate": "2022-02-12T13:04:59.927Z",
      "content": "<p><a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> wouldn't it have been better to wait with giving such advices after the end of the competition? For you probably not because then you wouldn't get so many votes. <br>\nIt is ridiculous that once you fight against a high scoring notebook published 8 days before the end of the competition (which I appreciate) and then, 2 days before the end of the competition, you come out with such a \"How to get to the silver zone\" manual.</p>",
      "rawMarkdown": "@vladvdv wouldn't it have been better to wait with giving such advices after the end of the competition? For you probably not because then you wouldn't get so many votes. \nIt is ridiculous that once you fight against a high scoring notebook published 8 days before the end of the competition (which I appreciate) and then, 2 days before the end of the competition, you come out with such a \"How to get to the silver zone\" manual.",
      "votes": 7,
      "replies": [
        {
          "id": 1686969,
          "postDate": "2022-02-12T13:25:45.417Z",
          "content": "<p><a href=\"https://www.kaggle.com/blankaf\" target=\"_blank\">@blankaf</a> <br>\nI understand what you want to say, and I stand my grounds regarding publishing high scoring notebooks in the last weeks of the competitions. Even more, I am against giving explicit details in forum threads.<br>\nThe difference between the later mentioned and this forum thread (a very big one) is that those notebooks were a fork, copy and submit ready to go solution that got you in the silver zone without even understand basics about machine learning or computer vision. This specific forum thread is discussing ideas and architecture principles. There is no recipe for success inside this forum thread that with 2 clicks teleports you to silver medal, and furthermore all the suggestions are lacking details (Adjusting augmentation, Adjusting learning rate to correspond with the optimizer characteristics and batch size, Adjusting Tracking parameters). <br>\nNo hyper-parameter is mentioned, no resolution, no learning rate, no tracking algorithm details, nothing. Furthermore, my score 2 secret sauce ingredients for sure are not even mentioned here and I will disclose them after the competition ends and regarding the specific questions asked here, you can see that I don't give answer that I consider to be to in-depht<br>\nSaying all that, I appreciate what you are trying to say and I am on the same side as yours but this forum thread is not against what we stand for.</p>",
          "rawMarkdown": "@blankaf \nI understand what you want to say, and I stand my grounds regarding publishing high scoring notebooks in the last weeks of the competitions. Even more, I am against giving explicit details in forum threads.\nThe difference between the later mentioned and this forum thread (a very big one) is that those notebooks were a fork, copy and submit ready to go solution that got you in the silver zone without even understand basics about machine learning or computer vision. This specific forum thread is discussing ideas and architecture principles. There is no recipe for success inside this forum thread that with 2 clicks teleports you to silver medal, and furthermore all the suggestions are lacking details (Adjusting augmentation, Adjusting learning rate to correspond with the optimizer characteristics and batch size, Adjusting Tracking parameters). \nNo hyper-parameter is mentioned, no resolution, no learning rate, no tracking algorithm details, nothing. Furthermore, my score 2 secret sauce ingredients for sure are not even mentioned here and I will disclose them after the competition ends and regarding the specific questions asked here, you can see that I don't give answer that I consider to be to in-depht\nSaying all that, I appreciate what you are trying to say and I am on the same side as yours but this forum thread is not against what we stand for.\n \n",
          "votes": 3
        },
        {
          "id": 1686997,
          "postDate": "2022-02-12T14:13:27.463Z",
          "content": "<p><a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> you explained me your stand in depth. <br>\nBut I still can't see here any acceptable reason for starting such a discussion thread 2 days before the end of the competition. <br>\nNo, this time we don't stand on the same side.</p>",
          "rawMarkdown": "@vladvdv you explained me your stand in depth. \nBut I still can't see here any acceptable reason for starting such a discussion thread 2 days before the end of the competition. \nNo, this time we don't stand on the same side."
        },
        {
          "id": 1687020,
          "postDate": "2022-02-12T14:41:19.943Z",
          "content": "<p>Passing tips close to the end of the competition ends up favoring those who have the highest number/fastest gpus</p>",
          "rawMarkdown": "Passing tips close to the end of the competition ends up favoring those who have the highest number/fastest gpus",
          "votes": 1
        },
        {
          "id": 1687035,
          "postDate": "2022-02-12T14:58:27.237Z",
          "content": "<p><a href=\"https://www.kaggle.com/robsonsan\" target=\"_blank\">@robsonsan</a> I agree. Furthermore I think that:<br>\nPassing ready to go inference kernels close to the end of the competition favors the beginners.<br>\nPassing explicit tips (\"use parameter x with value y\", \"use that resolution\") close to the end of the competition favors those who have the highest number/fastest gpus<br>\nDiscussing principles and architectures close to the end of the competition favors those who have a good know how of computer vision/machine learning. They can extrapolate from abstract concepts and ideas to possible solutions.<br>\n<a href=\"https://www.kaggle.com/blankaf\" target=\"_blank\">@blankaf</a> The purpose is to have a thread where we can all discuss high level ideas and concepts that we applied (as you can clearly see not explicit advice's or secret souses) for our last submissions. </p>",
          "rawMarkdown": "@robsonsan I agree. Furthermore I think that:\nPassing ready to go inference kernels close to the end of the competition favors the beginners.\nPassing explicit tips (\"use parameter x with value y\", \"use that resolution\") close to the end of the competition favors those who have the highest number/fastest gpus\nDiscussing principles and architectures close to the end of the competition favors those who have a good know how of computer vision/machine learning. They can extrapolate from abstract concepts and ideas to possible solutions.\n@blankaf The purpose is to have a thread where we can all discuss high level ideas and concepts that we applied (as you can clearly see not explicit advice's or secret souses) for our last submissions. "
        },
        {
          "id": 1687158,
          "postDate": "2022-02-12T16:40:58.900Z",
          "content": "<p>I really don't get what are you trying to achieve, shouting few days before for \"high scoring kernel\" and now this post, which seems like a clear recipe to me. <br>\nWhy you didn't wait 2 days more to share what worked and what not ?      </p>\n<blockquote>\n  <p>The purpose is to have a thread where we can all discuss high level ideas and concepts that we applied</p>\n</blockquote>\n<p>I don't see how </p>\n<blockquote>\n  <p>Adam instead of SGD</p>\n</blockquote>\n<p>promotes research on new ideas and learning..</p>\n<p>PS: I don't like downvoting posts, but rather express my disagreement  </p>",
          "rawMarkdown": "I really don't get what are you trying to achieve, shouting few days before for \"high scoring kernel\" and now this post, which seems like a clear recipe to me. \nWhy you didn't wait 2 days more to share what worked and what not ?      \n\n> The purpose is to have a thread where we can all discuss high level ideas and concepts that we applied\n\nI don't see how \n> Adam instead of SGD\n\npromotes research on new ideas and learning..\n\nPS: I don't like downvoting posts, but rather express my disagreement  ",
          "votes": 2
        },
        {
          "id": 1687187,
          "postDate": "2022-02-12T16:54:25.293Z",
          "content": "<p>This post seems like a clear recipe to you ??? If this seems the same for you than we have completly different perspectives. <br>\nHow can you compare this post with a ready to submit silver area notebooks ? This post only specifies common sense facts without any details. There isn't even one aspect specified here that can be considered as a ready to implement into a solution.</p>",
          "rawMarkdown": "This post seems like a clear recipe to you ??? If this seems the same for you than we have completly different perspectives. \nHow can you compare this post with a ready to submit silver area notebooks ? This post only specifies common sense facts without any details. There isn't even one aspect specified here that can be considered as a ready to implement into a solution.\n",
          "votes": 6
        },
        {
          "id": 1687537,
          "postDate": "2022-02-12T23:07:54.803Z",
          "content": "<p>This would have been a great thread to bounce ideas after the competition has ended. However, I agree with the others that sharing <strong>anything</strong> this close to the end of the competition is of questionable help to the community as a whole. It is not good that it might benefit one group over another (whoever that group may be).</p>\n<p>I get that the ideas given here are vague and not a specific recipe for anything. But vague ideas are not generally helpful, especially most people (including myself) are reluctant to debate these ideas in detail until after the competition closes. So what is the use of discussing such vague ideas at this time? If anything, it only upsets those who don't find these ideas helpful but are worried it may unfairly benefit others.</p>",
          "rawMarkdown": "This would have been a great thread to bounce ideas after the competition has ended. However, I agree with the others that sharing **anything** this close to the end of the competition is of questionable help to the community as a whole. It is not good that it might benefit one group over another (whoever that group may be).\n\nI get that the ideas given here are vague and not a specific recipe for anything. But vague ideas are not generally helpful, especially most people (including myself) are reluctant to debate these ideas in detail until after the competition closes. So what is the use of discussing such vague ideas at this time? If anything, it only upsets those who don't find these ideas helpful but are worried it may unfairly benefit others.",
          "votes": 2
        },
        {
          "id": 1687641,
          "postDate": "2022-02-13T03:05:01.453Z",
          "content": "<p>These will not bring negative effect on competition, because these points are not secrets and also without any details. The people with good hardware can easily find these points earlier than now. And the most points I also tried and almost agreed with it.On the contrary, the people without good hardware will be benefit from these, the can turn to the right way.</p>",
          "rawMarkdown": "These will not bring negative effect on competition, because these points are not secrets and also without any details. The people with good hardware can easily find these points earlier than now. And the most points I also tried and almost agreed with it.On the contrary, the people without good hardware will be benefit from these, the can turn to the right way.",
          "votes": 3
        },
        {
          "id": 1687653,
          "postDate": "2022-02-13T03:19:17.670Z",
          "content": "<p>Most points in the topic can be easily find in discussion zone, if kagglers spend some time on discussion topic. It is more like summary than release.</p>",
          "rawMarkdown": "Most points in the topic can be easily find in discussion zone, if kagglers spend some time on discussion topic. It is more like summary than release.",
          "votes": 3
        }
      ]
    },
    {
      "id": 1688316,
      "postDate": "2022-02-13T15:19:54.700Z",
      "content": "<p>Really great job <a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> </p>",
      "rawMarkdown": "Really great job @vladvdv ",
      "votes": 4,
      "replies": [
        {
          "id": 1688326,
          "postDate": "2022-02-13T15:23:02.943Z",
          "content": "<p>Thank you <a href=\"https://www.kaggle.com/danushkumarv\" target=\"_blank\">@danushkumarv</a> </p>",
          "rawMarkdown": "Thank you @danushkumarv ",
          "votes": 1
        }
      ]
    },
    {
      "id": 1689243,
      "postDate": "2022-02-14T05:17:39.443Z",
      "content": "<p>I also tried CLAHE as well as gamma correction and they didn't seem to help my model (though I tried that transformation on the full training set). After applying the transformation, the COTS definitely seemed to stand out more so it was a bit strange why the F2 score would drop. Do you have any theories on why CLAHE may not have been an effective augmentation with this dataset? I did notice that the landscape had a similar texture as the COTS (COTS was almost indistinguishable from the surroundings after sharpening the image twice).</p>",
      "rawMarkdown": "I also tried CLAHE as well as gamma correction and they didn't seem to help my model (though I tried that transformation on the full training set). After applying the transformation, the COTS definitely seemed to stand out more so it was a bit strange why the F2 score would drop. Do you have any theories on why CLAHE may not have been an effective augmentation with this dataset? I did notice that the landscape had a similar texture as the COTS (COTS was almost indistinguishable from the surroundings after sharpening the image twice).",
      "votes": 1
    },
    {
      "id": 1687925,
      "postDate": "2022-02-13T08:46:25.170Z",
      "content": "<p>thanks for sharing <a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> 😊</p>",
      "rawMarkdown": "thanks for sharing @vladvdv 😊",
      "votes": 1,
      "replies": [
        {
          "id": 1687932,
          "postDate": "2022-02-13T08:49:33.783Z",
          "content": "<p>With pleasure. Good luck in pursuing your learning path ! </p>",
          "rawMarkdown": "With pleasure. Good luck in pursuing your learning path ! "
        }
      ]
    },
    {
      "id": 1687010,
      "postDate": "2022-02-12T14:30:34.987Z",
      "content": "<p>Good advice!</p>",
      "rawMarkdown": "Good advice!",
      "votes": 1,
      "replies": [
        {
          "id": 1687017,
          "postDate": "2022-02-12T14:39:23.317Z",
          "content": "<p>Thanks, good luck ! </p>",
          "rawMarkdown": "Thanks, good luck ! ",
          "votes": 1
        }
      ]
    },
    {
      "id": 1690277,
      "postDate": "2022-02-14T20:36:43.230Z",
      "content": "<p>Nice summary <a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a>. We will talk in a moment about solutions … but have really simmilar experience. It is nice that you shared your observations. Thank you!  </p>",
      "rawMarkdown": "Nice summary @vladvdv. We will talk in a moment about solutions ... but have really simmilar experience. It is nice that you shared your observations. Thank you!  ",
      "votes": 2,
      "replies": [
        {
          "id": 1690292,
          "postDate": "2022-02-14T20:47:45.720Z",
          "content": "<p><a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> Thank you. Your notebooks and experience shared in the discussions were inspiring for us all. Great job and good luck tonight ! </p>",
          "rawMarkdown": "@remekkinas Thank you. Your notebooks and experience shared in the discussions were inspiring for us all. Great job and good luck tonight ! ",
          "votes": 1
        }
      ]
    },
    {
      "id": 1686881,
      "postDate": "2022-02-12T12:11:47.127Z",
      "content": "<p>looking forward to see the details after competition ends ! thx</p>",
      "rawMarkdown": "looking forward to see the details after competition ends ! thx",
      "votes": 2,
      "replies": [
        {
          "id": 1687024,
          "postDate": "2022-02-12T14:45:20.667Z",
          "content": "<p><a href=\"https://www.kaggle.com/simonsanghoonkim\" target=\"_blank\">@simonsanghoonkim</a>  Hopefully the shake will be kind to us. Good luck ! </p>",
          "rawMarkdown": "@simonsanghoonkim  Hopefully the shake will be kind to us. Good luck ! ",
          "votes": 1
        }
      ]
    },
    {
      "id": 1686873,
      "postDate": "2022-02-12T12:01:43.317Z",
      "content": "<p>Do you only use yolov5 or something else?</p>",
      "rawMarkdown": "Do you only use yolov5 or something else?",
      "replies": [
        {
          "id": 1686888,
          "postDate": "2022-02-12T12:15:35.640Z",
          "content": "<p><a href=\"https://www.kaggle.com/rishengyang\" target=\"_blank\">@rishengyang</a> Yes. My best solution is based on yoloV5</p>",
          "rawMarkdown": "@rishengyang Yes. My best solution is based on yoloV5"
        }
      ]
    },
    {
      "id": 1686747,
      "postDate": "2022-02-12T10:12:53.843Z",
      "content": "<p>Good luck and Thanks for sharing !</p>",
      "rawMarkdown": "Good luck and Thanks for sharing !\n",
      "replies": [
        {
          "id": 1686751,
          "postDate": "2022-02-12T10:14:48.423Z",
          "content": "<p>Good luck to you too <a href=\"https://www.kaggle.com/asalhi\" target=\"_blank\">@asalhi</a> !</p>",
          "rawMarkdown": "Good luck to you too @asalhi !",
          "votes": 1
        }
      ]
    },
    {
      "id": 1689597,
      "postDate": "2022-02-14T10:57:41.447Z",
      "content": "<p>Please explain to me what do you imply by \"Adjusting augmentation\"</p>",
      "rawMarkdown": "Please explain to me what do you imply by \"Adjusting augmentation\""
    },
    {
      "id": 1687427,
      "postDate": "2022-02-12T20:54:51.493Z",
      "content": "<p>By referring to \"whole data\", you retrained the model on all positive data or the whole 23k images provided? I'm currently adding the whole 23k images for retraining before final submisisons.</p>",
      "rawMarkdown": "By referring to \"whole data\", you retrained the model on all positive data or the whole 23k images provided? I'm currently adding the whole 23k images for retraining before final submisisons.",
      "replies": [
        {
          "id": 1689279,
          "postDate": "2022-02-14T06:12:16.733Z",
          "content": "<p>The answer has been displayed.<br>\n<strong>Ideas that did not work:</strong></p>\n<ul>\n<li>Adding no-label images to training data in a proportion of 5-10% of the total training data</li>\n</ul>",
          "rawMarkdown": "The answer has been displayed.\n**Ideas that did not work:**\n- Adding no-label images to training data in a proportion of 5-10% of the total training data\n"
        }
      ]
    },
    {
      "id": 1686726,
      "postDate": "2022-02-12T09:59:25.847Z",
      "rawMarkdown": "",
      "votes": 2,
      "isDeleted": true,
      "replies": [
        {
          "id": 1686737,
          "postDate": "2022-02-12T10:03:37.813Z",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/zzrysoseries\" target=\"_blank\">@zzrysoseries</a> I will talk in more details about my experience with tracking this after the competition ends, if my methodology will work of course</p>",
          "rawMarkdown": "Hi @zzrysoseries I will talk in more details about my experience with tracking this after the competition ends, if my methodology will work of course",
          "votes": 1
        },
        {
          "id": 1686741,
          "postDate": "2022-02-12T10:08:48.070Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1691186,
      "postDate": "2022-02-15T09:27:19.510Z",
      "content": "<p>Thanks for Sharing</p>",
      "rawMarkdown": "Thanks for Sharing"
    },
    {
      "id": 1690226,
      "postDate": "2022-02-14T19:44:57.347Z",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!"
    }
  ],
  "comments": [
    {
      "id": 1686942,
      "author_name": "Allie K.",
      "author_url": "",
      "post_date": "2022-02-12T13:04:59.927000",
      "content": "<p><a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> wouldn't it have been better to wait with giving such advices after the end of the competition? For you probably not because then you wouldn't get so many votes. <br>\nIt is ridiculous that once you fight against a high scoring notebook published 8 days before the end of the competition (which I appreciate) and then, 2 days before the end of the competition, you come out with such a \"How to get to the silver zone\" manual.</p>",
      "votes": 7,
      "replies": [
        {
          "id": 1686969,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T13:25:45.417000",
          "content": "<p><a href=\"https://www.kaggle.com/blankaf\" target=\"_blank\">@blankaf</a> <br>\nI understand what you want to say, and I stand my grounds regarding publishing high scoring notebooks in the last weeks of the competitions. Even more, I am against giving explicit details in forum threads.<br>\nThe difference between the later mentioned and this forum thread (a very big one) is that those notebooks were a fork, copy and submit ready to go solution that got you in the silver zone without even understand basics about machine learning or computer vision. This specific forum thread is discussing ideas and architecture principles. There is no recipe for success inside this forum thread that with 2 clicks teleports you to silver medal, and furthermore all the suggestions are lacking details (Adjusting augmentation, Adjusting learning rate to correspond with the optimizer characteristics and batch size, Adjusting Tracking parameters). <br>\nNo hyper-parameter is mentioned, no resolution, no learning rate, no tracking algorithm details, nothing. Furthermore, my score 2 secret sauce ingredients for sure are not even mentioned here and I will disclose them after the competition ends and regarding the specific questions asked here, you can see that I don't give answer that I consider to be to in-depht<br>\nSaying all that, I appreciate what you are trying to say and I am on the same side as yours but this forum thread is not against what we stand for.</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 1686997,
          "author_name": "Allie K.",
          "author_url": "",
          "post_date": "2022-02-12T14:13:27.463000",
          "content": "<p><a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> you explained me your stand in depth. <br>\nBut I still can't see here any acceptable reason for starting such a discussion thread 2 days before the end of the competition. <br>\nNo, this time we don't stand on the same side.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1687020,
          "author_name": "Robson",
          "author_url": "",
          "post_date": "2022-02-12T14:41:19.943000",
          "content": "<p>Passing tips close to the end of the competition ends up favoring those who have the highest number/fastest gpus</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1687035,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T14:58:27.237000",
          "content": "<p><a href=\"https://www.kaggle.com/robsonsan\" target=\"_blank\">@robsonsan</a> I agree. Furthermore I think that:<br>\nPassing ready to go inference kernels close to the end of the competition favors the beginners.<br>\nPassing explicit tips (\"use parameter x with value y\", \"use that resolution\") close to the end of the competition favors those who have the highest number/fastest gpus<br>\nDiscussing principles and architectures close to the end of the competition favors those who have a good know how of computer vision/machine learning. They can extrapolate from abstract concepts and ideas to possible solutions.<br>\n<a href=\"https://www.kaggle.com/blankaf\" target=\"_blank\">@blankaf</a> The purpose is to have a thread where we can all discuss high level ideas and concepts that we applied (as you can clearly see not explicit advice's or secret souses) for our last submissions. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1687158,
          "author_name": "Ioannis M",
          "author_url": "",
          "post_date": "2022-02-12T16:40:58.900000",
          "content": "<p>I really don't get what are you trying to achieve, shouting few days before for \"high scoring kernel\" and now this post, which seems like a clear recipe to me. <br>\nWhy you didn't wait 2 days more to share what worked and what not ?      </p>\n<blockquote>\n  <p>The purpose is to have a thread where we can all discuss high level ideas and concepts that we applied</p>\n</blockquote>\n<p>I don't see how </p>\n<blockquote>\n  <p>Adam instead of SGD</p>\n</blockquote>\n<p>promotes research on new ideas and learning..</p>\n<p>PS: I don't like downvoting posts, but rather express my disagreement  </p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1687187,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T16:54:25.293000",
          "content": "<p>This post seems like a clear recipe to you ??? If this seems the same for you than we have completly different perspectives. <br>\nHow can you compare this post with a ready to submit silver area notebooks ? This post only specifies common sense facts without any details. There isn't even one aspect specified here that can be considered as a ready to implement into a solution.</p>",
          "votes": 6,
          "replies": []
        },
        {
          "id": 1687537,
          "author_name": "Alex Wong",
          "author_url": "",
          "post_date": "2022-02-12T23:07:54.803000",
          "content": "<p>This would have been a great thread to bounce ideas after the competition has ended. However, I agree with the others that sharing <strong>anything</strong> this close to the end of the competition is of questionable help to the community as a whole. It is not good that it might benefit one group over another (whoever that group may be).</p>\n<p>I get that the ideas given here are vague and not a specific recipe for anything. But vague ideas are not generally helpful, especially most people (including myself) are reluctant to debate these ideas in detail until after the competition closes. So what is the use of discussing such vague ideas at this time? If anything, it only upsets those who don't find these ideas helpful but are worried it may unfairly benefit others.</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1687641,
          "author_name": "Good Moon",
          "author_url": "",
          "post_date": "2022-02-13T03:05:01.453000",
          "content": "<p>These will not bring negative effect on competition, because these points are not secrets and also without any details. The people with good hardware can easily find these points earlier than now. And the most points I also tried and almost agreed with it.On the contrary, the people without good hardware will be benefit from these, the can turn to the right way.</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 1687653,
          "author_name": "Good Moon",
          "author_url": "",
          "post_date": "2022-02-13T03:19:17.670000",
          "content": "<p>Most points in the topic can be easily find in discussion zone, if kagglers spend some time on discussion topic. It is more like summary than release.</p>",
          "votes": 3,
          "replies": []
        }
      ]
    },
    {
      "id": 1688316,
      "author_name": "Danushkumar Venkadesh",
      "author_url": "",
      "post_date": "2022-02-13T15:19:54.700000",
      "content": "<p>Really great job <a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> </p>",
      "votes": 4,
      "replies": [
        {
          "id": 1688326,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-13T15:23:02.943000",
          "content": "<p>Thank you <a href=\"https://www.kaggle.com/danushkumarv\" target=\"_blank\">@danushkumarv</a> </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1689243,
      "author_name": "Shreyas Agarwal",
      "author_url": "",
      "post_date": "2022-02-14T05:17:39.443000",
      "content": "<p>I also tried CLAHE as well as gamma correction and they didn't seem to help my model (though I tried that transformation on the full training set). After applying the transformation, the COTS definitely seemed to stand out more so it was a bit strange why the F2 score would drop. Do you have any theories on why CLAHE may not have been an effective augmentation with this dataset? I did notice that the landscape had a similar texture as the COTS (COTS was almost indistinguishable from the surroundings after sharpening the image twice).</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1687925,
      "author_name": "Aruna S",
      "author_url": "",
      "post_date": "2022-02-13T08:46:25.170000",
      "content": "<p>thanks for sharing <a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a> 😊</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1687932,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-13T08:49:33.783000",
          "content": "<p>With pleasure. Good luck in pursuing your learning path ! </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1687010,
      "author_name": "kaggler",
      "author_url": "",
      "post_date": "2022-02-12T14:30:34.987000",
      "content": "<p>Good advice!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1687017,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T14:39:23.317000",
          "content": "<p>Thanks, good luck ! </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1690277,
      "author_name": "Remek Kinas",
      "author_url": "",
      "post_date": "2022-02-14T20:36:43.230000",
      "content": "<p>Nice summary <a href=\"https://www.kaggle.com/vladvdv\" target=\"_blank\">@vladvdv</a>. We will talk in a moment about solutions … but have really simmilar experience. It is nice that you shared your observations. Thank you!  </p>",
      "votes": 2,
      "replies": [
        {
          "id": 1690292,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-14T20:47:45.720000",
          "content": "<p><a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> Thank you. Your notebooks and experience shared in the discussions were inspiring for us all. Great job and good luck tonight ! </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1686881,
      "author_name": "Sanghoon Kim",
      "author_url": "",
      "post_date": "2022-02-12T12:11:47.127000",
      "content": "<p>looking forward to see the details after competition ends ! thx</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1687024,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T14:45:20.667000",
          "content": "<p><a href=\"https://www.kaggle.com/simonsanghoonkim\" target=\"_blank\">@simonsanghoonkim</a>  Hopefully the shake will be kind to us. Good luck ! </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1686873,
      "author_name": "Kira yang",
      "author_url": "",
      "post_date": "2022-02-12T12:01:43.317000",
      "content": "<p>Do you only use yolov5 or something else?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1686888,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T12:15:35.640000",
          "content": "<p><a href=\"https://www.kaggle.com/rishengyang\" target=\"_blank\">@rishengyang</a> Yes. My best solution is based on yoloV5</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1686747,
      "author_name": "Ali",
      "author_url": "",
      "post_date": "2022-02-12T10:12:53.843000",
      "content": "<p>Good luck and Thanks for sharing !</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1686751,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T10:14:48.423000",
          "content": "<p>Good luck to you too <a href=\"https://www.kaggle.com/asalhi\" target=\"_blank\">@asalhi</a> !</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1689597,
      "author_name": "Muhammad Ammar Jamshed",
      "author_url": "",
      "post_date": "2022-02-14T10:57:41.447000",
      "content": "<p>Please explain to me what do you imply by \"Adjusting augmentation\"</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1687427,
      "author_name": "AnhNN",
      "author_url": "",
      "post_date": "2022-02-12T20:54:51.493000",
      "content": "<p>By referring to \"whole data\", you retrained the model on all positive data or the whole 23k images provided? I'm currently adding the whole 23k images for retraining before final submisisons.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1689279,
          "author_name": "Good Moon",
          "author_url": "",
          "post_date": "2022-02-14T06:12:16.733000",
          "content": "<p>The answer has been displayed.<br>\n<strong>Ideas that did not work:</strong></p>\n<ul>\n<li>Adding no-label images to training data in a proportion of 5-10% of the total training data</li>\n</ul>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1686726,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-02-12T09:59:25.847000",
      "content": "",
      "votes": 2,
      "replies": [
        {
          "id": 1686737,
          "author_name": "Vlad Vaduva",
          "author_url": "",
          "post_date": "2022-02-12T10:03:37.813000",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/zzrysoseries\" target=\"_blank\">@zzrysoseries</a> I will talk in more details about my experience with tracking this after the competition ends, if my methodology will work of course</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1686741,
          "author_name": "",
          "author_url": "",
          "post_date": "2022-02-12T10:08:48.070000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1691186,
      "author_name": "Arpita Shetty",
      "author_url": "",
      "post_date": "2022-02-15T09:27:19.510000",
      "content": "<p>Thanks for Sharing</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1690226,
      "author_name": "Abhinav Kumar",
      "author_url": "",
      "post_date": "2022-02-14T19:44:57.347000",
      "content": "<p>Thanks for sharing!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1686657": "**Ideas that worked:**\n- Using small architecture like Yolo5s6\n- Adam instead of SGD\n- Larger training image size than the default images sizes\n- Adjusting augmentation (Here is important to adapt to the model size, augmentations for Yolo5s6 are not enough for Yolo5l6 in order to generalize) \n- Retraining with whole data after selecting the proper model\n- Adjusting learning rate to correspond with the optimizer characteristics and batch size (wandb is very useful tool for observing the model evolution on training/valid)\n- Tracking (adjusting parameters may help more)\n\n**Ideas that did not work:**\n- Ensembles (so far). I am going to spend the remaining submissions optimizing ensembles techniques\n- Adding no-label images to training data in a proportion of 5-10% of the total training data\n- CLAHE\n- GAN techniques for increasing COTS in training data\n- SAHI  techniques did not worked for me, but I love the idea. Probably with more time invested in this, good results can be achieved\n- YOLOX got me lower performance comparative with YOLO5\n\nGood luck to everybody and congratulations to all participants for coming up with a tone of interesting ideas! This is one of the most interesting competitions from the last year ",
    "1686942": "@vladvdv wouldn't it have been better to wait with giving such advices after the end of the competition? For you probably not because then you wouldn't get so many votes. \nIt is ridiculous that once you fight against a high scoring notebook published 8 days before the end of the competition (which I appreciate) and then, 2 days before the end of the competition, you come out with such a \"How to get to the silver zone\" manual.",
    "1688316": "Really great job @vladvdv ",
    "1689243": "I also tried CLAHE as well as gamma correction and they didn't seem to help my model (though I tried that transformation on the full training set). After applying the transformation, the COTS definitely seemed to stand out more so it was a bit strange why the F2 score would drop. Do you have any theories on why CLAHE may not have been an effective augmentation with this dataset? I did notice that the landscape had a similar texture as the COTS (COTS was almost indistinguishable from the surroundings after sharpening the image twice).",
    "1687925": "thanks for sharing @vladvdv 😊",
    "1687010": "Good advice!",
    "1690277": "Nice summary @vladvdv. We will talk in a moment about solutions ... but have really simmilar experience. It is nice that you shared your observations. Thank you!  ",
    "1686881": "looking forward to see the details after competition ends ! thx",
    "1686873": "Do you only use yolov5 or something else?",
    "1686747": "Good luck and Thanks for sharing !\n",
    "1689597": "Please explain to me what do you imply by \"Adjusting augmentation\"",
    "1687427": "By referring to \"whole data\", you retrained the model on all positive data or the whole 23k images provided? I'm currently adding the whole 23k images for retraining before final submisisons.",
    "1686726": "",
    "1691186": "Thanks for Sharing",
    "1690226": "Thanks for sharing!"
  }
}