{
  "id": 99594,
  "title": "Data augmentation",
  "url": "/competitions/recursion-cellular-image-classification/discussion/99594",
  "author_name": "",
  "post_date": "2019-07-12T12:01:46.635823800Z",
  "votes": 4,
  "comment_count": 4,
  "views": 0,
  "content": "<p>What kind of data augmentation/transformation are you using? I apply standard  techniques like rotations, flips, shear and random crop (300x300), it does not give me a huge improvement (0.287 -&gt; 0.307). May be it is because I use relatively small model (resnet18) and it is not prone to overfit. </p>",
  "messages": [
    {
      "id": "573550",
      "postDate": "07/12/2019 12:01:46",
      "content": "<p>What kind of data augmentation/transformation are you using? I apply standard  techniques like rotations, flips, shear and random crop (300x300), it does not give me a huge improvement (0.287 -&gt; 0.307). May be it is because I use relatively small model (resnet18) and it is not prone to overfit. </p>",
      "rawMarkdown": "What kind of data augmentation/transformation are you using? I apply standard  techniques like rotations, flips, shear and random crop (300x300), it does not give me a huge improvement (0.287 -&gt; 0.307). May be it is because I use relatively small model (resnet18) and it is not prone to overfit.",
      "votes": null
    },
    {
      "id": "576794",
      "postDate": "07/16/2019 03:00:18",
      "content": "<p>Hi, may I know why did you choose to use resnet18? thanks!</p>",
      "rawMarkdown": "Hi, may I know why did you choose to use resnet18? thanks!",
      "votes": null
    },
    {
      "id": "578199",
      "postDate": "07/17/2019 13:08:52",
      "content": "<p>because it relatively small so I can prototype fast and check some ideas </p>",
      "rawMarkdown": "because it relatively small so I can prototype fast and check some ideas",
      "votes": null
    },
    {
      "id": "588054",
      "postDate": "07/30/2019 04:37:12",
      "content": "<p><a href=\"/cutlass90\">@cutlass90</a> Are you using PyTorch? If yes, are you using Lambda functions to apply the transformations to all the channels?</p>",
      "rawMarkdown": "cutlass90 Are you using PyTorch? If yes, are you using Lambda functions to apply the transformations to all the channels?",
      "votes": null
    },
    {
      "id": "588395",
      "postDate": "07/30/2019 14:15:36",
      "content": "<p>hi <a href=\"/cutlass90\">@cutlass90</a> is this a test time or train time augmentation you are doing? thanks! :) </p>",
      "rawMarkdown": "hi @cutlass90 is this a test time or train time augmentation you are doing? thanks! :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 576794,
      "author_name": "wjshenggggg",
      "author_url": "",
      "post_date": "07/16/2019 03:00:18",
      "content": "<p>Hi, may I know why did you choose to use resnet18? thanks!</p>",
      "votes": null,
      "replies": [
        {
          "id": 578199,
          "author_name": "cutlass90",
          "author_url": "",
          "post_date": "07/17/2019 13:08:52",
          "content": "<p>because it relatively small so I can prototype fast and check some ideas </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 588395,
          "author_name": "wjshenggggg",
          "author_url": "",
          "post_date": "07/30/2019 14:15:36",
          "content": "<p>hi <a href=\"/cutlass90\">@cutlass90</a> is this a test time or train time augmentation you are doing? thanks! :) </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 588054,
      "author_name": "lorenzofabbri92",
      "author_url": "",
      "post_date": "07/30/2019 04:37:12",
      "content": "<p><a href=\"/cutlass90\">@cutlass90</a> Are you using PyTorch? If yes, are you using Lambda functions to apply the transformations to all the channels?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "573550": "What kind of data augmentation/transformation are you using? I apply standard  techniques like rotations, flips, shear and random crop (300x300), it does not give me a huge improvement (0.287 -&gt; 0.307). May be it is because I use relatively small model (resnet18) and it is not prone to overfit.",
    "576794": "Hi, may I know why did you choose to use resnet18? thanks!",
    "578199": "because it relatively small so I can prototype fast and check some ideas",
    "588054": "cutlass90 Are you using PyTorch? If yes, are you using Lambda functions to apply the transformations to all the channels?",
    "588395": "hi @cutlass90 is this a test time or train time augmentation you are doing? thanks! :)"
  },
  "source": "meta"
}