{
  "id": 155928,
  "title": "Augmentation and Size Techniques",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/155928",
  "author_name": "",
  "post_date": "2020-06-03T15:22:48.499067700Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>What sort of augmentation techniques are you guys using. I have used the very basic ones and since the color of the spots is important Im not touching the saturation and all. What other techniques could be used to improve scores. Another thing is how big a role size is playing here. Im trying to do progressive resizing and currently I downloaded the 224 sized images. Is it worth downloading and resizing them to larger scales. I primarily use pytorch so I havent used tfrecords as well.</p>",
  "messages": [
    {
      "id": "872881",
      "postDate": "06/03/2020 15:22:48",
      "content": "<p>What sort of augmentation techniques are you guys using. I have used the very basic ones and since the color of the spots is important Im not touching the saturation and all. What other techniques could be used to improve scores. Another thing is how big a role size is playing here. Im trying to do progressive resizing and currently I downloaded the 224 sized images. Is it worth downloading and resizing them to larger scales. I primarily use pytorch so I havent used tfrecords as well.</p>",
      "rawMarkdown": "What sort of augmentation techniques are you guys using. I have used the very basic ones and since the color of the spots is important Im not touching the saturation and all. What other techniques could be used to improve scores. Another thing is how big a role size is playing here. Im trying to do progressive resizing and currently I downloaded the 224 sized images. Is it worth downloading and resizing them to larger scales. I primarily use pytorch so I havent used tfrecords as well.",
      "votes": null
    },
    {
      "id": "873904",
      "postDate": "06/04/2020 14:04:20",
      "content": "<p>Browsing the starter notebooks, I see that most have resized images to squares. Most of the stock images are rectangular. When I resized them to squares, I'm concerned that some of the lesions started to look quite distorted to me. Is square resizing a good idea? What are other's experiences?</p>",
      "rawMarkdown": "Browsing the starter notebooks, I see that most have resized images to squares. Most of the stock images are rectangular. When I resized them to squares, I'm concerned that some of the lesions started to look quite distorted to me. Is square resizing a good idea? What are other's experiences?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 873904,
      "author_name": "yeeseng",
      "author_url": "",
      "post_date": "06/04/2020 14:04:20",
      "content": "<p>Browsing the starter notebooks, I see that most have resized images to squares. Most of the stock images are rectangular. When I resized them to squares, I'm concerned that some of the lesions started to look quite distorted to me. Is square resizing a good idea? What are other's experiences?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "872881": "What sort of augmentation techniques are you guys using. I have used the very basic ones and since the color of the spots is important Im not touching the saturation and all. What other techniques could be used to improve scores. Another thing is how big a role size is playing here. Im trying to do progressive resizing and currently I downloaded the 224 sized images. Is it worth downloading and resizing them to larger scales. I primarily use pytorch so I havent used tfrecords as well.",
    "873904": "Browsing the starter notebooks, I see that most have resized images to squares. Most of the stock images are rectangular. When I resized them to squares, I'm concerned that some of the lesions started to look quite distorted to me. Is square resizing a good idea? What are other's experiences?"
  },
  "source": "meta"
}