{
  "id": 100651,
  "title": "Only pre trained?",
  "url": "/competitions/aptos2019-blindness-detection/discussion/100651",
  "author_name": "",
  "post_date": "2019-07-19T21:32:24.616133100Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I only see pre trained NN/weights in public kernels. Is anybody not using pre trained and obtaining a good score?</p>",
  "messages": [
    {
      "id": "580228",
      "postDate": "07/19/2019 21:32:24",
      "content": "<p>I only see pre trained NN/weights in public kernels. Is anybody not using pre trained and obtaining a good score?</p>",
      "rawMarkdown": "I only see pre trained NN/weights in public kernels. Is anybody not using pre trained and obtaining a good score?",
      "votes": null
    },
    {
      "id": "580349",
      "postDate": "07/20/2019 03:06:36",
      "content": "<p>Well pretrained on imagenet is available for everyone and since there are relatively few samples, transfer learning is preferred.\nAlso pretraining does a lot of feature extraction, its merely the classifier that can be replaced.\nIf there were more images, maybe 100k? I could see some people training, but it would probably be the people who have beefy machines to train locally.</p>",
      "rawMarkdown": "Well pretrained on imagenet is available for everyone and since there are relatively few samples, transfer learning is preferred.\nAlso pretraining does a lot of feature extraction, its merely the classifier that can be replaced.\nIf there were more images, maybe 100k? I could see some people training, but it would probably be the people who have beefy machines to train locally.",
      "votes": null
    },
    {
      "id": "581185",
      "postDate": "07/21/2019 15:07:27",
      "content": "<p>In most cases, you will use pre-trained models.  because in most cases you will not have enough data to get good results.</p>",
      "rawMarkdown": "In most cases, you will use pre-trained models.  because in most cases you will not have enough data to get good results.",
      "votes": null
    },
    {
      "id": "618995",
      "postDate": "09/05/2019 18:04:37",
      "content": "<p>hi, how do I access the pre trained weights with the internet off in my kernel? Is it already available on kaggle somehow?</p>",
      "rawMarkdown": "hi, how do I access the pre trained weights with the internet off in my kernel? Is it already available on kaggle somehow?",
      "votes": null
    },
    {
      "id": "619002",
      "postDate": "09/05/2019 18:20:19",
      "content": "<p>Yes, you can add pretrained weights from public Kaggle datasets. They are available for Pytorch and Keras</p>",
      "rawMarkdown": "Yes, you can add pretrained weights from public Kaggle datasets. They are available for Pytorch and Keras",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 580349,
      "author_name": "sidhanthholalkere",
      "author_url": "",
      "post_date": "07/20/2019 03:06:36",
      "content": "<p>Well pretrained on imagenet is available for everyone and since there are relatively few samples, transfer learning is preferred.\nAlso pretraining does a lot of feature extraction, its merely the classifier that can be replaced.\nIf there were more images, maybe 100k? I could see some people training, but it would probably be the people who have beefy machines to train locally.</p>",
      "votes": null,
      "replies": [
        {
          "id": 618995,
          "author_name": "aashishghosh",
          "author_url": "",
          "post_date": "09/05/2019 18:04:37",
          "content": "<p>hi, how do I access the pre trained weights with the internet off in my kernel? Is it already available on kaggle somehow?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 619002,
          "author_name": "demonplus",
          "author_url": "",
          "post_date": "09/05/2019 18:20:19",
          "content": "<p>Yes, you can add pretrained weights from public Kaggle datasets. They are available for Pytorch and Keras</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 581185,
      "author_name": "jmourad100",
      "author_url": "",
      "post_date": "07/21/2019 15:07:27",
      "content": "<p>In most cases, you will use pre-trained models.  because in most cases you will not have enough data to get good results.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "580228": "I only see pre trained NN/weights in public kernels. Is anybody not using pre trained and obtaining a good score?",
    "580349": "Well pretrained on imagenet is available for everyone and since there are relatively few samples, transfer learning is preferred.\nAlso pretraining does a lot of feature extraction, its merely the classifier that can be replaced.\nIf there were more images, maybe 100k? I could see some people training, but it would probably be the people who have beefy machines to train locally.",
    "581185": "In most cases, you will use pre-trained models.  because in most cases you will not have enough data to get good results.",
    "618995": "hi, how do I access the pre trained weights with the internet off in my kernel? Is it already available on kaggle somehow?",
    "619002": "Yes, you can add pretrained weights from public Kaggle datasets. They are available for Pytorch and Keras"
  },
  "source": "meta"
}