{
  "id": 105219,
  "title": "Cropping, coloring, flipping w/ 6-channels - right approach",
  "url": "/competitions/recursion-cellular-image-classification/discussion/105219",
  "author_name": "",
  "post_date": "2019-08-21T20:30:23.619450600Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Following the line of <a href=\"https://arxiv.org/abs/1812.01187\">bag of tricks</a>, when performing data augmentation (DA) for each RGB Image training instance, the same DA is applied for all three channels of the image since this is \"the same image\" and not yet transformed into a tensor (from what I understand).   </p>\n\n<p>What, then, should be the right approach for 6-channels? In my view, the same random-cropping/coloring/flipping/rotating should be performed for all 6 channels of a training instance to remain constant with the original trick - because \"not performing the same random crop across the 6 channels\" would, according to my intuition, make the training examples \"unstable\". I thus made modifications to the pytorch original transforms methods to be able to follow through on this idea. </p>\n\n<p>So I'd like to ask: am I understanding this right or there is another angle to this?</p>",
  "messages": [
    {
      "id": "604854",
      "postDate": "08/21/2019 20:30:23",
      "content": "<p>Following the line of <a href=\"https://arxiv.org/abs/1812.01187\">bag of tricks</a>, when performing data augmentation (DA) for each RGB Image training instance, the same DA is applied for all three channels of the image since this is \"the same image\" and not yet transformed into a tensor (from what I understand).   </p>\n\n<p>What, then, should be the right approach for 6-channels? In my view, the same random-cropping/coloring/flipping/rotating should be performed for all 6 channels of a training instance to remain constant with the original trick - because \"not performing the same random crop across the 6 channels\" would, according to my intuition, make the training examples \"unstable\". I thus made modifications to the pytorch original transforms methods to be able to follow through on this idea. </p>\n\n<p>So I'd like to ask: am I understanding this right or there is another angle to this?</p>",
      "rawMarkdown": "Following the line of [bag of tricks](https://arxiv.org/abs/1812.01187), when performing data augmentation (DA) for each RGB Image training instance, the same DA is applied for all three channels of the image since this is \"the same image\" and not yet transformed into a tensor (from what I understand).   \n\nWhat, then, should be the right approach for 6-channels? In my view, the same random-cropping/coloring/flipping/rotating should be performed for all 6 channels of a training instance to remain constant with the original trick - because \"not performing the same random crop across the 6 channels\" would, according to my intuition, make the training examples \"unstable\". I thus made modifications to the pytorch original transforms methods to be able to follow through on this idea. \n\nSo I'd like to ask: am I understanding this right or there is another angle to this?",
      "votes": null
    },
    {
      "id": "605115",
      "postDate": "08/22/2019 04:46:23",
      "content": "<p>you are right</p>",
      "rawMarkdown": "you are right",
      "votes": null
    },
    {
      "id": "608068",
      "postDate": "08/26/2019 09:31:17",
      "content": "<p>Are you still using the stats from the <em>Bag of Tricks</em> paper even with 6 channels?</p>",
      "rawMarkdown": "Are you still using the stats from the *Bag of Tricks* paper even with 6 channels?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 605115,
      "author_name": "christofhenkel",
      "author_url": "",
      "post_date": "08/22/2019 04:46:23",
      "content": "<p>you are right</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 608068,
      "author_name": "lorenzofabbri92",
      "author_url": "",
      "post_date": "08/26/2019 09:31:17",
      "content": "<p>Are you still using the stats from the <em>Bag of Tricks</em> paper even with 6 channels?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "604854": "Following the line of [bag of tricks](https://arxiv.org/abs/1812.01187), when performing data augmentation (DA) for each RGB Image training instance, the same DA is applied for all three channels of the image since this is \"the same image\" and not yet transformed into a tensor (from what I understand).   \n\nWhat, then, should be the right approach for 6-channels? In my view, the same random-cropping/coloring/flipping/rotating should be performed for all 6 channels of a training instance to remain constant with the original trick - because \"not performing the same random crop across the 6 channels\" would, according to my intuition, make the training examples \"unstable\". I thus made modifications to the pytorch original transforms methods to be able to follow through on this idea. \n\nSo I'd like to ask: am I understanding this right or there is another angle to this?",
    "605115": "you are right",
    "608068": "Are you still using the stats from the *Bag of Tricks* paper even with 6 channels?"
  },
  "source": "meta"
}