{
  "id": 430476,
  "title": "28th place solution",
  "url": "/competitions/google-research-identify-contrails-reduce-global-warming/discussion/430476",
  "author_name": "Bull",
  "post_date": "2023-08-10T00:48:48.668000",
  "votes": 22,
  "comment_count": 10,
  "views": 0,
  "content": "<p><strong>Overview</strong></p>\n<ul>\n<li><p>Emsemble 14models</p></li>\n<li><p>Model architecture: Unet, Unet++</p></li>\n<li><p>Encoder: resnest200e, resnest101e, efficientnet-b7</p></li>\n<li><p>The region that over 1/2 labelers considered contrail is GT, but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered contrail.I trained the model to predict these four labels simultaneously.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4472591%2Fe7023ad79299578a50b17c99fa6a4d71%2F.png?generation=1691633654216697&amp;alt=media\" alt=\"\"><br>\n(<a href=\"https://www.kaggle.com/code/welshonionman/dataset-ashcolor-4labels\" target=\"_blank\">dataset notebook</a>, <a href=\"https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels\" target=\"_blank\">training notebook</a> )</p></li>\n<li><p>pseudo labeling</p></li>\n</ul>\n<p><strong>Augmentation</strong></p>\n<ul>\n<li>A.RandomRotate90</li>\n<li>A.RandomBrightnessContrast</li>\n<li>A.CoarseDropout</li>\n<li>A.RandomGridShuffle</li>\n</ul>",
  "messages": [
    {
      "id": 2382697,
      "postDate": "2023-08-10T00:48:48.667Z",
      "content": "<p><strong>Overview</strong></p>\n<ul>\n<li><p>Emsemble 14models</p></li>\n<li><p>Model architecture: Unet, Unet++</p></li>\n<li><p>Encoder: resnest200e, resnest101e, efficientnet-b7</p></li>\n<li><p>The region that over 1/2 labelers considered contrail is GT, but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered contrail.I trained the model to predict these four labels simultaneously.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4472591%2Fe7023ad79299578a50b17c99fa6a4d71%2F.png?generation=1691633654216697&amp;alt=media\" alt=\"\"><br>\n(<a href=\"https://www.kaggle.com/code/welshonionman/dataset-ashcolor-4labels\" target=\"_blank\">dataset notebook</a>, <a href=\"https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels\" target=\"_blank\">training notebook</a> )</p></li>\n<li><p>pseudo labeling</p></li>\n</ul>\n<p><strong>Augmentation</strong></p>\n<ul>\n<li>A.RandomRotate90</li>\n<li>A.RandomBrightnessContrast</li>\n<li>A.CoarseDropout</li>\n<li>A.RandomGridShuffle</li>\n</ul>",
      "rawMarkdown": "**Overview**\n- Emsemble 14models\n- Model architecture: Unet, Unet++\n- Encoder: resnest200e, resnest101e, efficientnet-b7\n- The region that over 1/2 labelers considered contrail is GT, but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered contrail.I trained the model to predict these four labels simultaneously.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4472591%2Fe7023ad79299578a50b17c99fa6a4d71%2F.png?generation=1691633654216697&alt=media)\n([dataset notebook](https://www.kaggle.com/code/welshonionman/dataset-ashcolor-4labels), [training notebook](https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels) )\n\n- pseudo labeling\n\n**Augmentation**\n- A.RandomRotate90\n- A.RandomBrightnessContrast\n- A.CoarseDropout\n- A.RandomGridShuffle\n\n\n\n\n\n\n\n",
      "votes": 22
    },
    {
      "id": 2382908,
      "postDate": "2023-08-10T05:12:12.083Z",
      "content": "<p>\"but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered \"</p>\n<p>i suppose you are using softmax for multi class classification and the predicted class for submission is argmax(probability) ?</p>\n<p>or are you using \"probability for agree 2/4\" &gt; threshold for submission?</p>\n<p>or are you using sigmoid for each class or ordinal regression ????</p>",
      "rawMarkdown": "\"but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered \"\n\ni suppose you are using softmax for multi class classification and the predicted class for submission is argmax(probability) ?\n\nor are you using \"probability for agree 2/4\" > threshold for submission?\n\nor are you using sigmoid for each class or ordinal regression ????\n",
      "votes": 3,
      "replies": [
        {
          "id": 2383542,
          "postDate": "2023-08-10T12:35:22.367Z",
          "content": "<p>I trained the model as a multi-label problem, not a multi-class problem.<br>\nSo, the model outputs the areas that are considered contrail by labelers greater than 0, 1/4, 1/2, and 3/4; for submission, only the areas that are considered contrail by labelers greater than 1/2 are used.</p>\n<p>I share the training code..<br>\n<a href=\"https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels\" target=\"_blank\">https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels</a></p>",
          "rawMarkdown": "I trained the model as a multi-label problem, not a multi-class problem.\nSo, the model outputs the areas that are considered contrail by labelers greater than 0, 1/4, 1/2, and 3/4; for submission, only the areas that are considered contrail by labelers greater than 1/2 are used.\n\nI share the training code..\nhttps://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels",
          "votes": 1
        }
      ]
    },
    {
      "id": 2382821,
      "postDate": "2023-08-10T04:01:04.290Z",
      "content": "<p>Congratulations and thank you for sharing. What time did it take to train unet++?</p>",
      "rawMarkdown": "Congratulations and thank you for sharing. What time did it take to train unet++?\n",
      "votes": 1,
      "replies": [
        {
          "id": 2383550,
          "postDate": "2023-08-10T12:42:26.180Z",
          "content": "<p>Training time when using pseudo label was about 24h(50epochs).Using pseudo labeling takes very long time🙄</p>",
          "rawMarkdown": "Training time when using pseudo label was about 24h(50epochs).Using pseudo labeling takes very long time🙄"
        }
      ]
    },
    {
      "id": 2382703,
      "postDate": "2023-08-10T01:09:42.937Z",
      "content": "<p>Congratulations. What is your training image size?</p>",
      "rawMarkdown": "Congratulations. What is your training image size?",
      "votes": 1,
      "replies": [
        {
          "id": 2382741,
          "postDate": "2023-08-10T02:16:27.517Z",
          "content": "<p><a href=\"https://www.kaggle.com/ynhuhu\" target=\"_blank\">@ynhuhu</a> <br>\nTraining image size is 256.</p>",
          "rawMarkdown": "@ynhuhu \nTraining image size is 256.",
          "votes": 1
        }
      ]
    },
    {
      "id": 2382852,
      "postDate": "2023-08-10T04:27:25.830Z",
      "content": "<p>Congratulations for securing top position in the competition.<br>\nThanks for sharing the info.<br>\nDo you apply augmentation during training or inferejce.<br>\nI tried RandomGuassianBlur() but observed the performance degraded. <br>\nCan you also comment on GPU time. Did you have separate notebooks for training and inference. </p>",
      "rawMarkdown": "Congratulations for securing top position in the competition.\nThanks for sharing the info.\nDo you apply augmentation during training or inferejce.\nI tried RandomGuassianBlur() but observed the performance degraded. \nCan you also comment on GPU time. Did you have separate notebooks for training and inference. ",
      "votes": 2,
      "replies": [
        {
          "id": 2383548,
          "postDate": "2023-08-10T12:40:23.253Z",
          "content": "<p>I apply augmentation only during training, TTA isnt used.<br>\nIn discussions, there were some who said that augmentation is not effective at all. It's curious.</p>\n<p>I separate notebooks for training and inference. <br>\nTraining time when using pseudo labeling was about 24h(50epochs).<br>\nIt was very long🙄</p>",
          "rawMarkdown": "I apply augmentation only during training, TTA isnt used.\nIn discussions, there were some who said that augmentation is not effective at all. It's curious.\n\nI separate notebooks for training and inference. \nTraining time when using pseudo labeling was about 24h(50epochs).\nIt was very long🙄",
          "votes": 1
        }
      ]
    },
    {
      "id": 2394535,
      "postDate": "2023-08-17T02:05:24.983Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 2382714,
      "postDate": "2023-08-10T01:22:40.983Z",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 2382908,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2023-08-10T05:12:12.083000",
      "content": "<p>\"but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered \"</p>\n<p>i suppose you are using softmax for multi class classification and the predicted class for submission is argmax(probability) ?</p>\n<p>or are you using \"probability for agree 2/4\" &gt; threshold for submission?</p>\n<p>or are you using sigmoid for each class or ordinal regression ????</p>",
      "votes": 3,
      "replies": [
        {
          "id": 2383542,
          "author_name": "Bull",
          "author_url": "",
          "post_date": "2023-08-10T12:35:22.367000",
          "content": "<p>I trained the model as a multi-label problem, not a multi-class problem.<br>\nSo, the model outputs the areas that are considered contrail by labelers greater than 0, 1/4, 1/2, and 3/4; for submission, only the areas that are considered contrail by labelers greater than 1/2 are used.</p>\n<p>I share the training code..<br>\n<a href=\"https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels\" target=\"_blank\">https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels</a></p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 2382821,
      "author_name": "NA",
      "author_url": "",
      "post_date": "2023-08-10T04:01:04.290000",
      "content": "<p>Congratulations and thank you for sharing. What time did it take to train unet++?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2383550,
          "author_name": "Bull",
          "author_url": "",
          "post_date": "2023-08-10T12:42:26.180000",
          "content": "<p>Training time when using pseudo label was about 24h(50epochs).Using pseudo labeling takes very long time🙄</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2382703,
      "author_name": "ynhuhu",
      "author_url": "",
      "post_date": "2023-08-10T01:09:42.937000",
      "content": "<p>Congratulations. What is your training image size?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2382741,
          "author_name": "Bull",
          "author_url": "",
          "post_date": "2023-08-10T02:16:27.517000",
          "content": "<p><a href=\"https://www.kaggle.com/ynhuhu\" target=\"_blank\">@ynhuhu</a> <br>\nTraining image size is 256.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 2382852,
      "author_name": "C R Suthikshn Kumar",
      "author_url": "",
      "post_date": "2023-08-10T04:27:25.830000",
      "content": "<p>Congratulations for securing top position in the competition.<br>\nThanks for sharing the info.<br>\nDo you apply augmentation during training or inferejce.<br>\nI tried RandomGuassianBlur() but observed the performance degraded. <br>\nCan you also comment on GPU time. Did you have separate notebooks for training and inference. </p>",
      "votes": 2,
      "replies": [
        {
          "id": 2383548,
          "author_name": "Bull",
          "author_url": "",
          "post_date": "2023-08-10T12:40:23.253000",
          "content": "<p>I apply augmentation only during training, TTA isnt used.<br>\nIn discussions, there were some who said that augmentation is not effective at all. It's curious.</p>\n<p>I separate notebooks for training and inference. <br>\nTraining time when using pseudo labeling was about 24h(50epochs).<br>\nIt was very long🙄</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 2394535,
      "author_name": "",
      "author_url": "",
      "post_date": "2023-08-17T02:05:24.983000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2382714,
      "author_name": "po6ouw1",
      "author_url": "",
      "post_date": "2023-08-10T01:22:40.983000",
      "content": "<p>Thanks for sharing!</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2382697": "**Overview**\n- Emsemble 14models\n- Model architecture: Unet, Unet++\n- Encoder: resnest200e, resnest101e, efficientnet-b7\n- The region that over 1/2 labelers considered contrail is GT, but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered contrail.I trained the model to predict these four labels simultaneously.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4472591%2Fe7023ad79299578a50b17c99fa6a4d71%2F.png?generation=1691633654216697&alt=media)\n([dataset notebook](https://www.kaggle.com/code/welshonionman/dataset-ashcolor-4labels), [training notebook](https://www.kaggle.com/code/welshonionman/gr-icrgw-training-with-ashcolor-4labels) )\n\n- pseudo labeling\n\n**Augmentation**\n- A.RandomRotate90\n- A.RandomBrightnessContrast\n- A.CoarseDropout\n- A.RandomGridShuffle\n\n\n\n\n\n\n\n",
    "2382908": "\"but I created labels for the regions that over 0, 1/4, 1/2, and 3/4 labelers considered \"\n\ni suppose you are using softmax for multi class classification and the predicted class for submission is argmax(probability) ?\n\nor are you using \"probability for agree 2/4\" > threshold for submission?\n\nor are you using sigmoid for each class or ordinal regression ????\n",
    "2382821": "Congratulations and thank you for sharing. What time did it take to train unet++?\n",
    "2382703": "Congratulations. What is your training image size?",
    "2382852": "Congratulations for securing top position in the competition.\nThanks for sharing the info.\nDo you apply augmentation during training or inferejce.\nI tried RandomGuassianBlur() but observed the performance degraded. \nCan you also comment on GPU time. Did you have separate notebooks for training and inference. ",
    "2394535": "",
    "2382714": "Thanks for sharing!"
  }
}