{
  "id": 73423,
  "title": "Submission Format Question",
  "url": "/competitions/humpback-whale-identification/discussion/73423",
  "author_name": "Potter",
  "post_date": "2018-12-03T03:36:39.575000",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I'm having trouble understanding the submission format .csv. This is my first Kaggle competition...</p>\n\n<p>The sample_submission.csv shows two columns (Image, Id)\nAnd in the id it has 4 different whale ids repeated for every column.</p>\n\n<p>I think there are 5005 different whale ids in the challenge.</p>\n\n<p>Are we supposed to submit the 1 whale we think is most correct for each test image?\nAre we supposed to submit 5005 columns with the probability for each being the most correct?\nAre we supposed to submit the top 4 choices in the single Id column?</p>\n\n<p>Thank you,\n   David</p>",
  "messages": [
    {
      "id": 3065798,
      "postDate": "2024-12-07T10:04:47.433Z",
      "content": "<p>One aspect I find particularly intriguing is the potential use of transfer learning with pre-trained convolutional neural networks (CNNs) like ResNet or EfficientNet. These architectures might excel at capturing subtle details in whale flukes, like unique patterns and shapes. Has anyone experimented with fine-tuning such models specifically for this dataset?</p>\n<p>Additionally, data augmentation could be a game-changer here. Techniques like rotation, flipping, or even synthetic generation might help account for variations in lighting, angles, and occlusions. Any tips on balancing augmentation without overfitting?</p>\n<p>Excited to collaborate and see innovative solutions emerge from this community—let's push the boundaries of what AI can achieve in wildlife conservation!</p>\n<p>If this is helpful for you, please vote up to me!😄</p>",
      "rawMarkdown": "One aspect I find particularly intriguing is the potential use of transfer learning with pre-trained convolutional neural networks (CNNs) like ResNet or EfficientNet. These architectures might excel at capturing subtle details in whale flukes, like unique patterns and shapes. Has anyone experimented with fine-tuning such models specifically for this dataset?\n\nAdditionally, data augmentation could be a game-changer here. Techniques like rotation, flipping, or even synthetic generation might help account for variations in lighting, angles, and occlusions. Any tips on balancing augmentation without overfitting?\n\nExcited to collaborate and see innovative solutions emerge from this community—let's push the boundaries of what AI can achieve in wildlife conservation!\n\nIf this is helpful for you, please vote up to me!😄",
      "votes": 4
    },
    {
      "id": 431872,
      "postDate": "2018-12-03T03:36:39.577Z",
      "content": "<p>I'm having trouble understanding the submission format .csv. This is my first Kaggle competition...</p>\n\n<p>The sample_submission.csv shows two columns (Image, Id)\nAnd in the id it has 4 different whale ids repeated for every column.</p>\n\n<p>I think there are 5005 different whale ids in the challenge.</p>\n\n<p>Are we supposed to submit the 1 whale we think is most correct for each test image?\nAre we supposed to submit 5005 columns with the probability for each being the most correct?\nAre we supposed to submit the top 4 choices in the single Id column?</p>\n\n<p>Thank you,\n   David</p>",
      "rawMarkdown": "I'm having trouble understanding the submission format .csv. This is my first Kaggle competition...\n\nThe sample_submission.csv shows two columns (Image, Id)\nAnd in the id it has 4 different whale ids repeated for every column.\n\nI think there are 5005 different whale ids in the challenge.\n\nAre we supposed to submit the 1 whale we think is most correct for each test image?\nAre we supposed to submit 5005 columns with the probability for each being the most correct?\nAre we supposed to submit the top 4 choices in the single Id column?\n\nThank you,\n   David\n\n",
      "votes": 2
    },
    {
      "id": 431880,
      "postDate": "2018-12-03T03:53:20.070Z",
      "content": "<p>A common approach is to submit top-5 most probable classed for each test image.</p>",
      "rawMarkdown": "A common approach is to submit top-5 most probable classed for each test image."
    },
    {
      "id": 431926,
      "postDate": "2018-12-03T05:25:22.740Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 431888,
      "postDate": "2018-12-03T04:02:38.817Z",
      "content": "<p>Makes sense. Thanks Andrew</p>",
      "rawMarkdown": "Makes sense. Thanks Andrew"
    }
  ],
  "comments": [
    {
      "id": 3065798,
      "author_name": "Kevin Smith",
      "author_url": "",
      "post_date": "2024-12-07T10:04:47.433000",
      "content": "<p>One aspect I find particularly intriguing is the potential use of transfer learning with pre-trained convolutional neural networks (CNNs) like ResNet or EfficientNet. These architectures might excel at capturing subtle details in whale flukes, like unique patterns and shapes. Has anyone experimented with fine-tuning such models specifically for this dataset?</p>\n<p>Additionally, data augmentation could be a game-changer here. Techniques like rotation, flipping, or even synthetic generation might help account for variations in lighting, angles, and occlusions. Any tips on balancing augmentation without overfitting?</p>\n<p>Excited to collaborate and see innovative solutions emerge from this community—let's push the boundaries of what AI can achieve in wildlife conservation!</p>\n<p>If this is helpful for you, please vote up to me!😄</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 431880,
      "author_name": "Andrey Lukyanenko",
      "author_url": "",
      "post_date": "2018-12-03T03:53:20.070000",
      "content": "<p>A common approach is to submit top-5 most probable classed for each test image.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 431926,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-12-03T05:25:22.740000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 431888,
      "author_name": "Potter",
      "author_url": "",
      "post_date": "2018-12-03T04:02:38.817000",
      "content": "<p>Makes sense. Thanks Andrew</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3065798": "One aspect I find particularly intriguing is the potential use of transfer learning with pre-trained convolutional neural networks (CNNs) like ResNet or EfficientNet. These architectures might excel at capturing subtle details in whale flukes, like unique patterns and shapes. Has anyone experimented with fine-tuning such models specifically for this dataset?\n\nAdditionally, data augmentation could be a game-changer here. Techniques like rotation, flipping, or even synthetic generation might help account for variations in lighting, angles, and occlusions. Any tips on balancing augmentation without overfitting?\n\nExcited to collaborate and see innovative solutions emerge from this community—let's push the boundaries of what AI can achieve in wildlife conservation!\n\nIf this is helpful for you, please vote up to me!😄",
    "431872": "I'm having trouble understanding the submission format .csv. This is my first Kaggle competition...\n\nThe sample_submission.csv shows two columns (Image, Id)\nAnd in the id it has 4 different whale ids repeated for every column.\n\nI think there are 5005 different whale ids in the challenge.\n\nAre we supposed to submit the 1 whale we think is most correct for each test image?\nAre we supposed to submit 5005 columns with the probability for each being the most correct?\nAre we supposed to submit the top 4 choices in the single Id column?\n\nThank you,\n   David\n\n",
    "431880": "A common approach is to submit top-5 most probable classed for each test image.",
    "431926": "",
    "431888": "Makes sense. Thanks Andrew"
  }
}