{
  "id": 169988,
  "title": "Start Small, Go Big",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/169988",
  "author_name": "Signal",
  "post_date": "2020-07-26T03:03:41.031000",
  "votes": 5,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Take your model of choice: Resnet50, EfficientNet-B2, etc.  Train it on given image size, say 128x128.  Now take that same model, and instead feed it 192x192 (basically a bigger size).  you tweak it.  Now feed it 256x256.  You will see an improvement.  Why does this work? Who knows, but it does and saves you some time.  Sometimes the benefits are better than training the model from scratch on the new image size.  </p>",
  "messages": [
    {
      "id": 945617,
      "postDate": "2020-07-26T03:03:41.033Z",
      "content": "<p>Take your model of choice: Resnet50, EfficientNet-B2, etc.  Train it on given image size, say 128x128.  Now take that same model, and instead feed it 192x192 (basically a bigger size).  you tweak it.  Now feed it 256x256.  You will see an improvement.  Why does this work? Who knows, but it does and saves you some time.  Sometimes the benefits are better than training the model from scratch on the new image size.  </p>",
      "rawMarkdown": "Take your model of choice: Resnet50, EfficientNet-B2, etc.  Train it on given image size, say 128x128.  Now take that same model, and instead feed it 192x192 (basically a bigger size).  you tweak it.  Now feed it 256x256.  You will see an improvement.  Why does this work? Who knows, but it does and saves you some time.  Sometimes the benefits are better than training the model from scratch on the new image size.  ",
      "votes": 4
    },
    {
      "id": 947186,
      "postDate": "2020-07-27T05:53:51.120Z",
      "content": "<p>Do you do transfer learning from image size to image size? Seems line an interesting approach. What kind of improvements are we talking about?</p>",
      "rawMarkdown": "Do you do transfer learning from image size to image size? Seems line an interesting approach. What kind of improvements are we talking about?",
      "replies": [
        {
          "id": 947446,
          "postDate": "2020-07-27T09:20:38.177Z",
          "content": "<p>Yeah you can think of it like that, do some experiments and see what it does for you</p>",
          "rawMarkdown": "Yeah you can think of it like that, do some experiments and see what it does for you"
        }
      ]
    },
    {
      "id": 946226,
      "postDate": "2020-07-26T12:53:12.033Z",
      "content": "<p>Do you mean CV improvement or LB improvement or both ? Thanks.</p>",
      "rawMarkdown": "Do you mean CV improvement or LB improvement or both ? Thanks.",
      "replies": [
        {
          "id": 947390,
          "postDate": "2020-07-27T08:35:54.790Z",
          "content": "<p>It should effect both.  This is not specific to this contest this is just a general tip for Transfer Learning.  You can expect to squeeze out a percent or two.</p>",
          "rawMarkdown": "It should effect both.  This is not specific to this contest this is just a general tip for Transfer Learning.  You can expect to squeeze out a percent or two.",
          "votes": 2
        }
      ]
    },
    {
      "id": 946235,
      "postDate": "2020-07-26T13:01:22.503Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 947186,
      "author_name": "André Neves",
      "author_url": "",
      "post_date": "2020-07-27T05:53:51.120000",
      "content": "<p>Do you do transfer learning from image size to image size? Seems line an interesting approach. What kind of improvements are we talking about?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 947446,
          "author_name": "Signal",
          "author_url": "",
          "post_date": "2020-07-27T09:20:38.177000",
          "content": "<p>Yeah you can think of it like that, do some experiments and see what it does for you</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 946226,
      "author_name": "Changyi",
      "author_url": "",
      "post_date": "2020-07-26T12:53:12.033000",
      "content": "<p>Do you mean CV improvement or LB improvement or both ? Thanks.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 947390,
          "author_name": "Signal",
          "author_url": "",
          "post_date": "2020-07-27T08:35:54.790000",
          "content": "<p>It should effect both.  This is not specific to this contest this is just a general tip for Transfer Learning.  You can expect to squeeze out a percent or two.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 946235,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-07-26T13:01:22.503000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "945617": "Take your model of choice: Resnet50, EfficientNet-B2, etc.  Train it on given image size, say 128x128.  Now take that same model, and instead feed it 192x192 (basically a bigger size).  you tweak it.  Now feed it 256x256.  You will see an improvement.  Why does this work? Who knows, but it does and saves you some time.  Sometimes the benefits are better than training the model from scratch on the new image size.  ",
    "947186": "Do you do transfer learning from image size to image size? Seems line an interesting approach. What kind of improvements are we talking about?",
    "946226": "Do you mean CV improvement or LB improvement or both ? Thanks.",
    "946235": ""
  }
}