{
  "id": 99656,
  "title": "transfer learning",
  "url": "/competitions/generative-dog-images/discussion/99656",
  "author_name": "",
  "post_date": "2019-07-13T02:38:21.304294700Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Is transfer learning allowed to cut down training time or do you have to train from scratch?</p>",
  "messages": [
    {
      "id": "573951",
      "postDate": "07/13/2019 02:38:21",
      "content": "<p>Is transfer learning allowed to cut down training time or do you have to train from scratch?</p>",
      "rawMarkdown": "Is transfer learning allowed to cut down training time or do you have to train from scratch?",
      "votes": null
    },
    {
      "id": "574187",
      "postDate": "07/13/2019 10:48:27",
      "content": "<p>In this competition, your kernel is not allowed to connect to the internet nor load any datasets. So you cannot load any publicly pre-trained nets. However within your kernel, you can train multiple nets and transfer learning between them.</p>",
      "rawMarkdown": "In this competition, your kernel is not allowed to connect to the internet nor load any datasets. So you cannot load any publicly pre-trained nets. However within your kernel, you can train multiple nets and transfer learning between them.",
      "votes": null
    },
    {
      "id": "574887",
      "postDate": "07/14/2019 16:16:56",
      "content": "<p>yeah, I figured that would be the case, although i think to promote best pratices that are followed in industry and academia, competitions like this should allow transfer learning. Training from scratch just seems like a waste of time and electricity and your result will be better is far less time with transfer learning.</p>",
      "rawMarkdown": "yeah, I figured that would be the case, although i think to promote best pratices that are followed in industry and academia, competitions like this should allow transfer learning. Training from scratch just seems like a waste of time and electricity and your result will be better is far less time with transfer learning.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 574187,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "07/13/2019 10:48:27",
      "content": "<p>In this competition, your kernel is not allowed to connect to the internet nor load any datasets. So you cannot load any publicly pre-trained nets. However within your kernel, you can train multiple nets and transfer learning between them.</p>",
      "votes": null,
      "replies": [
        {
          "id": 574887,
          "author_name": "ahsenk",
          "author_url": "",
          "post_date": "07/14/2019 16:16:56",
          "content": "<p>yeah, I figured that would be the case, although i think to promote best pratices that are followed in industry and academia, competitions like this should allow transfer learning. Training from scratch just seems like a waste of time and electricity and your result will be better is far less time with transfer learning.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "573951": "Is transfer learning allowed to cut down training time or do you have to train from scratch?",
    "574187": "In this competition, your kernel is not allowed to connect to the internet nor load any datasets. So you cannot load any publicly pre-trained nets. However within your kernel, you can train multiple nets and transfer learning between them.",
    "574887": "yeah, I figured that would be the case, although i think to promote best pratices that are followed in industry and academia, competitions like this should allow transfer learning. Training from scratch just seems like a waste of time and electricity and your result will be better is far less time with transfer learning."
  },
  "source": "meta"
}