{
  "id": 319829,
  "title": "6th place solution",
  "url": "/competitions/happy-whale-and-dolphin/discussion/319829",
  "author_name": "YujiAriyasu",
  "post_date": "2022-04-19T04:11:43.810000",
  "votes": 57,
  "comment_count": 22,
  "views": 0,
  "content": "<p>Congrats to all the winners.<br>\nThanks to Kaggle and the hosting team for an interesting competition.<br>\nI look forward to the third one:)</p>\n<p>Here is my solution summary.</p>\n<h1>Dataset</h1>\n<p>I used the full body and back fin data created by Jan. And I also used the results of training the detector using Jan's annotations. there were two different boxes for each fullbody / backfin.<br>\nI also used data with a slightly larger box. As a result, a fairly diverse set of data was used as material for the ensemble. And full images were also used as material for the ensemble too.<br>\nImage sizes: 512 ~ 896.</p>\n<h1>validation</h1>\n<p>Since I started working with tensorflow late and was short on time, I took the approach of training with all trains and checking scores in LB.<br>\nDevelopment without oof was quite difficult, but this time I thought it would not be a problem because the shake seemed to be quite small.</p>\n<h1>Model</h1>\n<h3>tensorflow</h3>\n<p>All models are connected to dolg and arcface.<br>\nDynamic margins were equally accurate with or without. Both were used.</p>\n<ul>\n<li>efficientnet v1: 5 / 6 / 7 / l2</li>\n<li>efficientnet v2: l / xl</li>\n<li>convnext: l / xl<br>\nSince I started using tensorflow in April, I ended up using the augmentation and hyperparameters as they are in the public notebook.</li>\n</ul>\n<h3>pytorch</h3>\n<p>All models are connected to arcface.(without dolg, without dynamic margins)</p>\n<ul>\n<li>convnext :xl</li>\n<li>efficientnet: l2</li>\n<li>swintransformer: large384 (image size was 768)<br>\nI used a fairly heavy augmentation.</li>\n</ul>\n<h1>Inference</h1>\n<p>I compared the similarity of the concated feature map between train and test.<br>\nThe dimension of the final feature map exceeded 20,000.<br>\nDifferent thresholds were used to determine new individual id for each species.<br>\nno pp.</p>\n<h1>iterative pseudo labeling</h1>\n<p>By using pseudo labeling, I can see not only the train but also the similarity to the confident test set. This is why pseudo labeling is important in this competition. So by repeating pseudo labeling multiple times, I was able to improve the score little by little.</p>\n<h1>cat cafe</h1>\n<p>Since this was an 'animal competition', working in a cat cafe greatly improved my score.<br>\n<img src=\"https://user-images.githubusercontent.com/28746788/163918410-ae3f65e9-f620-4006-abb5-403a448e73f3.png\" alt=\"cat.png\"></p>",
  "messages": [
    {
      "id": 1760101,
      "postDate": "2022-04-19T04:11:43.810Z",
      "content": "<p>Congrats to all the winners.<br>\nThanks to Kaggle and the hosting team for an interesting competition.<br>\nI look forward to the third one:)</p>\n<p>Here is my solution summary.</p>\n<h1>Dataset</h1>\n<p>I used the full body and back fin data created by Jan. And I also used the results of training the detector using Jan's annotations. there were two different boxes for each fullbody / backfin.<br>\nI also used data with a slightly larger box. As a result, a fairly diverse set of data was used as material for the ensemble. And full images were also used as material for the ensemble too.<br>\nImage sizes: 512 ~ 896.</p>\n<h1>validation</h1>\n<p>Since I started working with tensorflow late and was short on time, I took the approach of training with all trains and checking scores in LB.<br>\nDevelopment without oof was quite difficult, but this time I thought it would not be a problem because the shake seemed to be quite small.</p>\n<h1>Model</h1>\n<h3>tensorflow</h3>\n<p>All models are connected to dolg and arcface.<br>\nDynamic margins were equally accurate with or without. Both were used.</p>\n<ul>\n<li>efficientnet v1: 5 / 6 / 7 / l2</li>\n<li>efficientnet v2: l / xl</li>\n<li>convnext: l / xl<br>\nSince I started using tensorflow in April, I ended up using the augmentation and hyperparameters as they are in the public notebook.</li>\n</ul>\n<h3>pytorch</h3>\n<p>All models are connected to arcface.(without dolg, without dynamic margins)</p>\n<ul>\n<li>convnext :xl</li>\n<li>efficientnet: l2</li>\n<li>swintransformer: large384 (image size was 768)<br>\nI used a fairly heavy augmentation.</li>\n</ul>\n<h1>Inference</h1>\n<p>I compared the similarity of the concated feature map between train and test.<br>\nThe dimension of the final feature map exceeded 20,000.<br>\nDifferent thresholds were used to determine new individual id for each species.<br>\nno pp.</p>\n<h1>iterative pseudo labeling</h1>\n<p>By using pseudo labeling, I can see not only the train but also the similarity to the confident test set. This is why pseudo labeling is important in this competition. So by repeating pseudo labeling multiple times, I was able to improve the score little by little.</p>\n<h1>cat cafe</h1>\n<p>Since this was an 'animal competition', working in a cat cafe greatly improved my score.<br>\n<img src=\"https://user-images.githubusercontent.com/28746788/163918410-ae3f65e9-f620-4006-abb5-403a448e73f3.png\" alt=\"cat.png\"></p>",
      "rawMarkdown": "Congrats to all the winners.\nThanks to Kaggle and the hosting team for an interesting competition.\nI look forward to the third one:)\n\nHere is my solution summary.\n\n# Dataset\nI used the full body and back fin data created by Jan. And I also used the results of training the detector using Jan's annotations. there were two different boxes for each fullbody / backfin.\nI also used data with a slightly larger box. As a result, a fairly diverse set of data was used as material for the ensemble. And full images were also used as material for the ensemble too.\nImage sizes: 512 ~ 896.\n\n# validation\nSince I started working with tensorflow late and was short on time, I took the approach of training with all trains and checking scores in LB.\nDevelopment without oof was quite difficult, but this time I thought it would not be a problem because the shake seemed to be quite small.\n\n# Model\n### tensorflow\nAll models are connected to dolg and arcface.\nDynamic margins were equally accurate with or without. Both were used.\n- efficientnet v1: 5 / 6 / 7 / l2\n- efficientnet v2: l / xl\n- convnext: l / xl\nSince I started using tensorflow in April, I ended up using the augmentation and hyperparameters as they are in the public notebook.\n\n### pytorch\nAll models are connected to arcface.(without dolg, without dynamic margins)\n- convnext :xl\n- efficientnet: l2\n- swintransformer: large384 (image size was 768)\nI used a fairly heavy augmentation.\n\n# Inference\nI compared the similarity of the concated feature map between train and test.\nThe dimension of the final feature map exceeded 20,000.\nDifferent thresholds were used to determine new individual id for each species.\nno pp.\n\n# iterative pseudo labeling\nBy using pseudo labeling, I can see not only the train but also the similarity to the confident test set. This is why pseudo labeling is important in this competition. So by repeating pseudo labeling multiple times, I was able to improve the score little by little.\n\n# cat cafe\nSince this was an 'animal competition', working in a cat cafe greatly improved my score.\n![cat.png](https://user-images.githubusercontent.com/28746788/163918410-ae3f65e9-f620-4006-abb5-403a448e73f3.png)\n",
      "votes": 57
    },
    {
      "id": 1760275,
      "postDate": "2022-04-19T07:11:42.670Z",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/yujiariyasu\" target=\"_blank\">@yujiariyasu</a> on results. One more to gold :)</p>",
      "rawMarkdown": "Congrats @yujiariyasu on results. One more to gold :)",
      "votes": 1,
      "replies": [
        {
          "id": 1760350,
          "postDate": "2022-04-19T08:26:14Z",
          "content": "<p>Thank you:)</p>",
          "rawMarkdown": "Thank you:)"
        }
      ]
    },
    {
      "id": 1760130,
      "postDate": "2022-04-19T04:53:15.373Z",
      "content": "<p>Hi! Interesting sum-up.<br>\nSeems like you upload HUGE amount of data in kaggle datasets, for example on 15.04.2022 you make 154 datasets versions!<br>\nCan you tell us, what this data are and did it helps to improve you score?</p>",
      "rawMarkdown": "Hi! Interesting sum-up.\nSeems like you upload HUGE amount of data in kaggle datasets, for example on 15.04.2022 you make 154 datasets versions!\nCan you tell us, what this data are and did it helps to improve you score?",
      "votes": 1,
      "replies": [
        {
          "id": 1760166,
          "postDate": "2022-04-19T05:25:04.697Z",
          "content": "<p>Because of the large size, I divided the data into 7 kaggle datasets per type of data. The variations are two types of fullbody and backfin data each, and multiple types of box sizes. To save training time as much as possible, I created pre-cropped data. Of course, this contributed greatly to the accuracy.</p>",
          "rawMarkdown": "Because of the large size, I divided the data into 7 kaggle datasets per type of data. The variations are two types of fullbody and backfin data each, and multiple types of box sizes. To save training time as much as possible, I created pre-cropped data. Of course, this contributed greatly to the accuracy.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1760110,
      "postDate": "2022-04-19T04:26:37.397Z",
      "content": "<p>Can you explain how you create your pseudo label ?</p>",
      "rawMarkdown": "Can you explain how you create your pseudo label ?",
      "votes": 1,
      "replies": [
        {
          "id": 1760121,
          "postDate": "2022-04-19T04:38:18.190Z",
          "content": "<p>\"Different thresholds were used to determine new individual id for each species.\" and about this, you trained another model for species classification right ? </p>",
          "rawMarkdown": "\"Different thresholds were used to determine new individual id for each species.\" and about this, you trained another model for species classification right ? "
        },
        {
          "id": 1760164,
          "postDate": "2022-04-19T05:22:31.213Z",
          "content": "<p>Yes, And I treated test which is close to train as a new train.</p>",
          "rawMarkdown": "Yes, And I treated test which is close to train as a new train."
        },
        {
          "id": 1760168,
          "postDate": "2022-04-19T05:25:52.793Z",
          "content": "<p>Any note about how you choose high quality pseudo label? score above threshold or just use highest one?</p>",
          "rawMarkdown": "Any note about how you choose high quality pseudo label? score above threshold or just use highest one?"
        },
        {
          "id": 1760174,
          "postDate": "2022-04-19T05:37:17.680Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1760183,
          "postDate": "2022-04-19T05:51:11.603Z",
          "content": "<p>score above threshold. Finally, about 2,0000 more were added.</p>",
          "rawMarkdown": "score above threshold. Finally, about 2,0000 more were added."
        }
      ]
    },
    {
      "id": 1770550,
      "postDate": "2022-04-28T11:10:30.983Z",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/yujiariyasu\" target=\"_blank\">@yujiariyasu</a> for your solution ! </p>\n<p>What was your hardware to train such models ?</p>\n<p>If working in a cat coffee boosts your score, I guess if a HappyCat competition is to be released soon, you'll have to take more risks to boost your score 😄</p>\n<p>Congrats again </p>",
      "rawMarkdown": "Congrats @yujiariyasu for your solution ! \n\nWhat was your hardware to train such models ?\n\nIf working in a cat coffee boosts your score, I guess if a HappyCat competition is to be released soon, you'll have to take more risks to boost your score 😄\n\nCongrats again ",
      "replies": [
        {
          "id": 1770594,
          "postDate": "2022-04-28T12:05:19.423Z",
          "content": "<p>Thanks!! 🐈<br>\nI used multiple colab pro+ and kaggle TPUs. 🐈🐈</p>",
          "rawMarkdown": "Thanks!! 🐈\nI used multiple colab pro+ and kaggle TPUs. 🐈🐈",
          "votes": 1
        },
        {
          "id": 1770615,
          "postDate": "2022-04-28T12:27:43.300Z",
          "content": "<p>Amazing ! Gotta try pro+, still a bit limited with pro in terms of memory<br>\nThanks sir</p>",
          "rawMarkdown": "Amazing ! Gotta try pro+, still a bit limited with pro in terms of memory\nThanks sir"
        }
      ]
    },
    {
      "id": 1761618,
      "postDate": "2022-04-20T03:27:49.223Z",
      "content": "<p>Congratulations to our new GM candidate!! 🙌</p>",
      "rawMarkdown": "Congratulations to our new GM candidate!! 🙌",
      "replies": [
        {
          "id": 1761634,
          "postDate": "2022-04-20T03:47:28.393Z",
          "content": "<p>Thanks, Qishen!<br>\nGM is just a passing point, but I am still happy that I am so close to achieving it.<br>\nI will try my best to be like you one day!</p>",
          "rawMarkdown": "Thanks, Qishen!\nGM is just a passing point, but I am still happy that I am so close to achieving it.\nI will try my best to be like you one day!"
        }
      ]
    },
    {
      "id": 1760348,
      "postDate": "2022-04-19T08:23:19.140Z",
      "content": "<p>Congratulations! Let me ask you a few questions.</p>\n<ol>\n<li>I also tried DOLG with TensorFlow but it did not improve my score. I used <a href=\"https://github.com/innat/DOLG-TensorFlow\" target=\"_blank\">this</a> as a reference for my implementation, did you have any tips on implementing DOLG?</li>\n<li>how much did your score increase when you changed the threshold for each species?</li>\n</ol>",
      "rawMarkdown": "Congratulations! Let me ask you a few questions.\n1. I also tried DOLG with TensorFlow but it did not improve my score. I used [this](https://github.com/innat/DOLG-TensorFlow) as a reference for my implementation, did you have any tips on implementing DOLG?\n2. how much did your score increase when you changed the threshold for each species?",
      "replies": [
        {
          "id": 1760352,
          "postDate": "2022-04-19T08:28:23.410Z",
          "content": "<p>1, The dieter implementation is helpful.<br>\n2, About 0.5%~1%.</p>",
          "rawMarkdown": "1, The dieter implementation is helpful.\n2, About 0.5%~1%.",
          "votes": 1
        },
        {
          "id": 1760372,
          "postDate": "2022-04-19T08:40:34.223Z",
          "content": "<p>Thank you for answering.</p>",
          "rawMarkdown": "Thank you for answering."
        }
      ]
    },
    {
      "id": 1760159,
      "postDate": "2022-04-19T05:19:42.330Z",
      "content": "<p>it's good to see DoLG model in top 10. i was not able to make it even though I used DoLG :P. if possible please explain your model implementation and tweaks for Eff v2 training. Thanks </p>",
      "rawMarkdown": "it's good to see DoLG model in top 10. i was not able to make it even though I used DoLG :P. if possible please explain your model implementation and tweaks for Eff v2 training. Thanks ",
      "replies": [
        {
          "id": 1760382,
          "postDate": "2022-04-19T08:50:18.187Z",
          "content": "<p>The dieter implementation is helpful.</p>",
          "rawMarkdown": "The dieter implementation is helpful.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1760105,
      "postDate": "2022-04-19T04:14:47.263Z",
      "content": "<p>Upvoted because cats</p>",
      "rawMarkdown": "Upvoted because cats",
      "replies": [
        {
          "id": 1760160,
          "postDate": "2022-04-19T05:20:24.857Z",
          "content": "<p>Thanks to you and cats:)</p>",
          "rawMarkdown": "Thanks to you and cats:)",
          "votes": 1
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1760275,
      "author_name": "KhanhVD",
      "author_url": "",
      "post_date": "2022-04-19T07:11:42.670000",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/yujiariyasu\" target=\"_blank\">@yujiariyasu</a> on results. One more to gold :)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1760350,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-19T08:26:14",
          "content": "<p>Thank you:)</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1760130,
      "author_name": "Ilya Dobrynin",
      "author_url": "",
      "post_date": "2022-04-19T04:53:15.373000",
      "content": "<p>Hi! Interesting sum-up.<br>\nSeems like you upload HUGE amount of data in kaggle datasets, for example on 15.04.2022 you make 154 datasets versions!<br>\nCan you tell us, what this data are and did it helps to improve you score?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1760166,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-19T05:25:04.697000",
          "content": "<p>Because of the large size, I divided the data into 7 kaggle datasets per type of data. The variations are two types of fullbody and backfin data each, and multiple types of box sizes. To save training time as much as possible, I created pre-cropped data. Of course, this contributed greatly to the accuracy.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1760110,
      "author_name": "cuongnn",
      "author_url": "",
      "post_date": "2022-04-19T04:26:37.397000",
      "content": "<p>Can you explain how you create your pseudo label ?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1760121,
          "author_name": "cuongnn",
          "author_url": "",
          "post_date": "2022-04-19T04:38:18.190000",
          "content": "<p>\"Different thresholds were used to determine new individual id for each species.\" and about this, you trained another model for species classification right ? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1760164,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-19T05:22:31.213000",
          "content": "<p>Yes, And I treated test which is close to train as a new train.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1760168,
          "author_name": "cuongnn",
          "author_url": "",
          "post_date": "2022-04-19T05:25:52.793000",
          "content": "<p>Any note about how you choose high quality pseudo label? score above threshold or just use highest one?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1760174,
          "author_name": "",
          "author_url": "",
          "post_date": "2022-04-19T05:37:17.680000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1760183,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-19T05:51:11.603000",
          "content": "<p>score above threshold. Finally, about 2,0000 more were added.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1770550,
      "author_name": "BryanB",
      "author_url": "",
      "post_date": "2022-04-28T11:10:30.983000",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/yujiariyasu\" target=\"_blank\">@yujiariyasu</a> for your solution ! </p>\n<p>What was your hardware to train such models ?</p>\n<p>If working in a cat coffee boosts your score, I guess if a HappyCat competition is to be released soon, you'll have to take more risks to boost your score 😄</p>\n<p>Congrats again </p>",
      "votes": 0,
      "replies": [
        {
          "id": 1770594,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-28T12:05:19.423000",
          "content": "<p>Thanks!! 🐈<br>\nI used multiple colab pro+ and kaggle TPUs. 🐈🐈</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1770615,
          "author_name": "BryanB",
          "author_url": "",
          "post_date": "2022-04-28T12:27:43.300000",
          "content": "<p>Amazing ! Gotta try pro+, still a bit limited with pro in terms of memory<br>\nThanks sir</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1761618,
      "author_name": "Qishen Ha",
      "author_url": "",
      "post_date": "2022-04-20T03:27:49.223000",
      "content": "<p>Congratulations to our new GM candidate!! 🙌</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1761634,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-20T03:47:28.393000",
          "content": "<p>Thanks, Qishen!<br>\nGM is just a passing point, but I am still happy that I am so close to achieving it.<br>\nI will try my best to be like you one day!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1760348,
      "author_name": "kurokami",
      "author_url": "",
      "post_date": "2022-04-19T08:23:19.140000",
      "content": "<p>Congratulations! Let me ask you a few questions.</p>\n<ol>\n<li>I also tried DOLG with TensorFlow but it did not improve my score. I used <a href=\"https://github.com/innat/DOLG-TensorFlow\" target=\"_blank\">this</a> as a reference for my implementation, did you have any tips on implementing DOLG?</li>\n<li>how much did your score increase when you changed the threshold for each species?</li>\n</ol>",
      "votes": 0,
      "replies": [
        {
          "id": 1760352,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-19T08:28:23.410000",
          "content": "<p>1, The dieter implementation is helpful.<br>\n2, About 0.5%~1%.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1760372,
          "author_name": "kurokami",
          "author_url": "",
          "post_date": "2022-04-19T08:40:34.223000",
          "content": "<p>Thank you for answering.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1760159,
      "author_name": "yuvaramsingh",
      "author_url": "",
      "post_date": "2022-04-19T05:19:42.330000",
      "content": "<p>it's good to see DoLG model in top 10. i was not able to make it even though I used DoLG :P. if possible please explain your model implementation and tweaks for Eff v2 training. Thanks </p>",
      "votes": 0,
      "replies": [
        {
          "id": 1760382,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-19T08:50:18.187000",
          "content": "<p>The dieter implementation is helpful.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1760105,
      "author_name": "outwrest",
      "author_url": "",
      "post_date": "2022-04-19T04:14:47.263000",
      "content": "<p>Upvoted because cats</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1760160,
          "author_name": "YujiAriyasu",
          "author_url": "",
          "post_date": "2022-04-19T05:20:24.857000",
          "content": "<p>Thanks to you and cats:)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1760101": "Congrats to all the winners.\nThanks to Kaggle and the hosting team for an interesting competition.\nI look forward to the third one:)\n\nHere is my solution summary.\n\n# Dataset\nI used the full body and back fin data created by Jan. And I also used the results of training the detector using Jan's annotations. there were two different boxes for each fullbody / backfin.\nI also used data with a slightly larger box. As a result, a fairly diverse set of data was used as material for the ensemble. And full images were also used as material for the ensemble too.\nImage sizes: 512 ~ 896.\n\n# validation\nSince I started working with tensorflow late and was short on time, I took the approach of training with all trains and checking scores in LB.\nDevelopment without oof was quite difficult, but this time I thought it would not be a problem because the shake seemed to be quite small.\n\n# Model\n### tensorflow\nAll models are connected to dolg and arcface.\nDynamic margins were equally accurate with or without. Both were used.\n- efficientnet v1: 5 / 6 / 7 / l2\n- efficientnet v2: l / xl\n- convnext: l / xl\nSince I started using tensorflow in April, I ended up using the augmentation and hyperparameters as they are in the public notebook.\n\n### pytorch\nAll models are connected to arcface.(without dolg, without dynamic margins)\n- convnext :xl\n- efficientnet: l2\n- swintransformer: large384 (image size was 768)\nI used a fairly heavy augmentation.\n\n# Inference\nI compared the similarity of the concated feature map between train and test.\nThe dimension of the final feature map exceeded 20,000.\nDifferent thresholds were used to determine new individual id for each species.\nno pp.\n\n# iterative pseudo labeling\nBy using pseudo labeling, I can see not only the train but also the similarity to the confident test set. This is why pseudo labeling is important in this competition. So by repeating pseudo labeling multiple times, I was able to improve the score little by little.\n\n# cat cafe\nSince this was an 'animal competition', working in a cat cafe greatly improved my score.\n![cat.png](https://user-images.githubusercontent.com/28746788/163918410-ae3f65e9-f620-4006-abb5-403a448e73f3.png)\n",
    "1760275": "Congrats @yujiariyasu on results. One more to gold :)",
    "1760130": "Hi! Interesting sum-up.\nSeems like you upload HUGE amount of data in kaggle datasets, for example on 15.04.2022 you make 154 datasets versions!\nCan you tell us, what this data are and did it helps to improve you score?",
    "1760110": "Can you explain how you create your pseudo label ?",
    "1770550": "Congrats @yujiariyasu for your solution ! \n\nWhat was your hardware to train such models ?\n\nIf working in a cat coffee boosts your score, I guess if a HappyCat competition is to be released soon, you'll have to take more risks to boost your score 😄\n\nCongrats again ",
    "1761618": "Congratulations to our new GM candidate!! 🙌",
    "1760348": "Congratulations! Let me ask you a few questions.\n1. I also tried DOLG with TensorFlow but it did not improve my score. I used [this](https://github.com/innat/DOLG-TensorFlow) as a reference for my implementation, did you have any tips on implementing DOLG?\n2. how much did your score increase when you changed the threshold for each species?",
    "1760159": "it's good to see DoLG model in top 10. i was not able to make it even though I used DoLG :P. if possible please explain your model implementation and tweaks for Eff v2 training. Thanks ",
    "1760105": "Upvoted because cats"
  }
}