{
  "id": 430481,
  "title": "[PB 0.685] Single Model Single Fold Imagesize 384",
  "url": "/competitions/google-research-identify-contrails-reduce-global-warming/discussion/430481",
  "author_name": "Chainey",
  "post_date": "2023-08-10T01:45:12.844000",
  "votes": 11,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I took part in this competition in the last week. <br>\nThanks for sharing your findings on the discussion board, I'm learning a lot.</p>\n<p>This model achieved 0.685 in my validation fold and 0.670 in the LB. I merged other models at the end, resulting in a decrease in the PB, 0.681.</p>\n<p><strong>Overview</strong></p>\n<p>Model architecture: Unet<br>\nEncoder: timm-efficientnet-b8</p>\n<p><strong>Train</strong></p>\n<p><strong>First Stage</strong>:</p>\n<p>Imagesize: 384<br>\nEpoch: 60<br>\nScheduler: LR 0.0005 CosineAnnealingLR T_max 60 eta_min 1.0e-6 weight_decay 0.0001<br>\nLoss: Dice<br>\nAugmentation (pseudo code):<br>\nCompose{<br>\nOneof{<br>\nNo change()<br>\nHF()<br>\nVF()<br>\nRot90(factor=2)<br>\n}<br>\nRandomBrightnessContrast()<br>\nRandomResizedCrop(256,(0.75, 1.0),0.5)<br>\n}</p>\n<p><strong>Second Stage</strong>:</p>\n<p>Pre-trained: weights(First stage)<br>\nImagesize: 384<br>\nEpoch: 50<br>\nScheduler: LR 0.0002 CosineAnnealingLR T_max 50 eta_min 1.0e-6 weight_decay 0.0001<br>\nLoss: Dice<br>\nAugmentation (pseudo code):<br>\nCompose{<br>\nHF()<br>\nVF()<br>\nRot90()<br>\nRandomResizedCrop(256,(0.75, 1.0),0.6)<br>\n}</p>\n<p><strong>Infer</strong></p>\n<p>Imagesize: 416<br>\nTTA: Flip and rotate combinations for a total of 8 different orientations(As many orientations are used in second stage training as in infering)</p>",
  "messages": [
    {
      "id": 2382726,
      "postDate": "2023-08-10T01:45:12.843Z",
      "content": "<p>I took part in this competition in the last week. <br>\nThanks for sharing your findings on the discussion board, I'm learning a lot.</p>\n<p>This model achieved 0.685 in my validation fold and 0.670 in the LB. I merged other models at the end, resulting in a decrease in the PB, 0.681.</p>\n<p><strong>Overview</strong></p>\n<p>Model architecture: Unet<br>\nEncoder: timm-efficientnet-b8</p>\n<p><strong>Train</strong></p>\n<p><strong>First Stage</strong>:</p>\n<p>Imagesize: 384<br>\nEpoch: 60<br>\nScheduler: LR 0.0005 CosineAnnealingLR T_max 60 eta_min 1.0e-6 weight_decay 0.0001<br>\nLoss: Dice<br>\nAugmentation (pseudo code):<br>\nCompose{<br>\nOneof{<br>\nNo change()<br>\nHF()<br>\nVF()<br>\nRot90(factor=2)<br>\n}<br>\nRandomBrightnessContrast()<br>\nRandomResizedCrop(256,(0.75, 1.0),0.5)<br>\n}</p>\n<p><strong>Second Stage</strong>:</p>\n<p>Pre-trained: weights(First stage)<br>\nImagesize: 384<br>\nEpoch: 50<br>\nScheduler: LR 0.0002 CosineAnnealingLR T_max 50 eta_min 1.0e-6 weight_decay 0.0001<br>\nLoss: Dice<br>\nAugmentation (pseudo code):<br>\nCompose{<br>\nHF()<br>\nVF()<br>\nRot90()<br>\nRandomResizedCrop(256,(0.75, 1.0),0.6)<br>\n}</p>\n<p><strong>Infer</strong></p>\n<p>Imagesize: 416<br>\nTTA: Flip and rotate combinations for a total of 8 different orientations(As many orientations are used in second stage training as in infering)</p>",
      "rawMarkdown": "I took part in this competition in the last week. \nThanks for sharing your findings on the discussion board, I'm learning a lot.\n\nThis model achieved 0.685 in my validation fold and 0.670 in the LB. I merged other models at the end, resulting in a decrease in the PB, 0.681.\n\n**Overview**\n\nModel architecture: Unet\nEncoder: timm-efficientnet-b8\n\n**Train**\n\n**First Stage**:\n\nImagesize: 384\nEpoch: 60\nScheduler: LR 0.0005 CosineAnnealingLR T_max 60 eta_min 1.0e-6 weight_decay 0.0001\nLoss: Dice\nAugmentation (pseudo code):\nCompose{\nOneof{\nNo change()\nHF()\nVF()\nRot90(factor=2)\n}\nRandomBrightnessContrast()\nRandomResizedCrop(256,(0.75, 1.0),0.5)\n}\n\n**Second Stage**:\n\nPre-trained: weights(First stage)\nImagesize: 384\nEpoch: 50\nScheduler: LR 0.0002 CosineAnnealingLR T_max 50 eta_min 1.0e-6 weight_decay 0.0001\nLoss: Dice\nAugmentation (pseudo code):\nCompose{\nHF()\nVF()\nRot90()\nRandomResizedCrop(256,(0.75, 1.0),0.6)\n}\n\n\n**Infer**\n\nImagesize: 416\nTTA: Flip and rotate combinations for a total of 8 different orientations(As many orientations are used in second stage training as in infering)\n\n",
      "votes": 10
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2382726": "I took part in this competition in the last week. \nThanks for sharing your findings on the discussion board, I'm learning a lot.\n\nThis model achieved 0.685 in my validation fold and 0.670 in the LB. I merged other models at the end, resulting in a decrease in the PB, 0.681.\n\n**Overview**\n\nModel architecture: Unet\nEncoder: timm-efficientnet-b8\n\n**Train**\n\n**First Stage**:\n\nImagesize: 384\nEpoch: 60\nScheduler: LR 0.0005 CosineAnnealingLR T_max 60 eta_min 1.0e-6 weight_decay 0.0001\nLoss: Dice\nAugmentation (pseudo code):\nCompose{\nOneof{\nNo change()\nHF()\nVF()\nRot90(factor=2)\n}\nRandomBrightnessContrast()\nRandomResizedCrop(256,(0.75, 1.0),0.5)\n}\n\n**Second Stage**:\n\nPre-trained: weights(First stage)\nImagesize: 384\nEpoch: 50\nScheduler: LR 0.0002 CosineAnnealingLR T_max 50 eta_min 1.0e-6 weight_decay 0.0001\nLoss: Dice\nAugmentation (pseudo code):\nCompose{\nHF()\nVF()\nRot90()\nRandomResizedCrop(256,(0.75, 1.0),0.6)\n}\n\n\n**Infer**\n\nImagesize: 416\nTTA: Flip and rotate combinations for a total of 8 different orientations(As many orientations are used in second stage training as in infering)\n\n"
  }
}