{
  "id": 539759,
  "title": "Training Data seems somewhat sparse",
  "url": "/competitions/eclipse-megamovie/discussion/539759",
  "author_name": "",
  "post_date": "2024-10-10T16:58:36.105376Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Just a quick look at the training data provided makes me want to change my mind on using a CNN to solve this challenge. There are not many samples of each partial eclipse to train accurately, in my opinion. I'll give it a try, but maybe some hybrid approach of determining the percentage of eclipse coverage will be better for that part of the challenge.</p>",
  "messages": [
    {
      "id": "3013899",
      "postDate": "10/10/2024 16:58:36",
      "content": "<p>Just a quick look at the training data provided makes me want to change my mind on using a CNN to solve this challenge. There are not many samples of each partial eclipse to train accurately, in my opinion. I'll give it a try, but maybe some hybrid approach of determining the percentage of eclipse coverage will be better for that part of the challenge.</p>",
      "rawMarkdown": "Just a quick look at the training data provided makes me want to change my mind on using a CNN to solve this challenge. There are not many samples of each partial eclipse to train accurately, in my opinion. I'll give it a try, but maybe some hybrid approach of determining the percentage of eclipse coverage will be better for that part of the challenge.",
      "votes": null
    },
    {
      "id": "3015466",
      "postDate": "10/12/2024 12:46:50",
      "content": "<p>I am facing the same problem. I have tried to train CNNs and other variants and I only get to 70% accuracy in the training. Not an easy one.</p>",
      "rawMarkdown": "I am facing the same problem. I have tried to train CNNs and other variants and I only get to 70% accuracy in the training. Not an easy one.",
      "votes": null
    },
    {
      "id": "3020129",
      "postDate": "10/17/2024 07:25:11",
      "content": "<p>For sure, around 500 images to train a model is not a lot. Possible approaches include preprocessing, data augmentation, and transfer learning using models pretrained on image classification. Good luck :)</p>",
      "rawMarkdown": "For sure, around 500 images to train a model is not a lot. Possible approaches include preprocessing, data augmentation, and transfer learning using models pretrained on image classification. Good luck :)",
      "votes": null
    },
    {
      "id": "3032691",
      "postDate": "10/31/2024 08:29:53",
      "content": "<p>I suspect meta learning might improve accuracy even further</p>",
      "rawMarkdown": "I suspect meta learning might improve accuracy even further",
      "votes": null
    },
    {
      "id": "3033371",
      "postDate": "11/01/2024 03:38:08",
      "content": "<p>Thanks, I might have to investigate meta learning. My attempt to use transfer learning on a VGG16 model looked OK for training accuracy, but only slightly better that 50% when run on the test images.</p>",
      "rawMarkdown": "Thanks, I might have to investigate meta learning. My attempt to use transfer learning on a VGG16 model looked OK for training accuracy, but only slightly better that 50% when run on the test images.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3015466,
      "author_name": "laguila11",
      "author_url": "",
      "post_date": "10/12/2024 12:46:50",
      "content": "<p>I am facing the same problem. I have tried to train CNNs and other variants and I only get to 70% accuracy in the training. Not an easy one.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3020129,
      "author_name": "clementhinderer",
      "author_url": "",
      "post_date": "10/17/2024 07:25:11",
      "content": "<p>For sure, around 500 images to train a model is not a lot. Possible approaches include preprocessing, data augmentation, and transfer learning using models pretrained on image classification. Good luck :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3032691,
      "author_name": "ucas0v0zhuoqunli",
      "author_url": "",
      "post_date": "10/31/2024 08:29:53",
      "content": "<p>I suspect meta learning might improve accuracy even further</p>",
      "votes": null,
      "replies": [
        {
          "id": 3033371,
          "author_name": "poolsidebill",
          "author_url": "",
          "post_date": "11/01/2024 03:38:08",
          "content": "<p>Thanks, I might have to investigate meta learning. My attempt to use transfer learning on a VGG16 model looked OK for training accuracy, but only slightly better that 50% when run on the test images.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3013899": "Just a quick look at the training data provided makes me want to change my mind on using a CNN to solve this challenge. There are not many samples of each partial eclipse to train accurately, in my opinion. I'll give it a try, but maybe some hybrid approach of determining the percentage of eclipse coverage will be better for that part of the challenge.",
    "3015466": "I am facing the same problem. I have tried to train CNNs and other variants and I only get to 70% accuracy in the training. Not an easy one.",
    "3020129": "For sure, around 500 images to train a model is not a lot. Possible approaches include preprocessing, data augmentation, and transfer learning using models pretrained on image classification. Good luck :)",
    "3032691": "I suspect meta learning might improve accuracy even further",
    "3033371": "Thanks, I might have to investigate meta learning. My attempt to use transfer learning on a VGG16 model looked OK for training accuracy, but only slightly better that 50% when run on the test images."
  },
  "source": "meta"
}