{
  "id": 199150,
  "title": "Question: Image sizing for Efficientnet ",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/199150",
  "author_name": "",
  "post_date": "2020-11-24T16:17:00.610679900Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Love all the great notebooks using Efficientnet but I had a question. I noticed most crop or resize these images to something like 300x300 or 512x512 when the images are 800x600. </p>\n<p>Thoughts on losing data when doing this? Or just using the same aspect ratio to scale these down? I'm not as familiar with Efficientnet so wanted to get some thoughts. Thanks!<br>\nCharlie</p>",
  "messages": [
    {
      "id": "1089597",
      "postDate": "11/24/2020 16:17:00",
      "content": "<p>Love all the great notebooks using Efficientnet but I had a question. I noticed most crop or resize these images to something like 300x300 or 512x512 when the images are 800x600. </p>\n<p>Thoughts on losing data when doing this? Or just using the same aspect ratio to scale these down? I'm not as familiar with Efficientnet so wanted to get some thoughts. Thanks!<br>\nCharlie</p>",
      "rawMarkdown": "Love all the great notebooks using Efficientnet but I had a question. I noticed most crop or resize these images to something like 300x300 or 512x512 when the images are 800x600. \n\nThoughts on losing data when doing this? Or just using the same aspect ratio to scale these down? I'm not as familiar with Efficientnet so wanted to get some thoughts. Thanks!\nCharlie",
      "votes": null
    },
    {
      "id": "1089847",
      "postDate": "11/24/2020 20:44:22",
      "content": "<p>The largest efficientnet (B7) was trained on 512x512 images, which is why a lot of people are resizing to that shape. Moreover, it was trained on resized images from ImageNet and was able to achieve 85%+ accuracy, so it's safe to say that the performance loss from resizing is negligible.</p>",
      "rawMarkdown": "The largest efficientnet (B7) was trained on 512x512 images, which is why a lot of people are resizing to that shape. Moreover, it was trained on resized images from ImageNet and was able to achieve 85%+ accuracy, so it's safe to say that the performance loss from resizing is negligible.",
      "votes": null
    },
    {
      "id": "1090463",
      "postDate": "11/25/2020 10:54:13",
      "content": "<p>Thank you. I appreciate the response!</p>",
      "rawMarkdown": "Thank you. I appreciate the response!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1089847,
      "author_name": "xhlulu",
      "author_url": "",
      "post_date": "11/24/2020 20:44:22",
      "content": "<p>The largest efficientnet (B7) was trained on 512x512 images, which is why a lot of people are resizing to that shape. Moreover, it was trained on resized images from ImageNet and was able to achieve 85%+ accuracy, so it's safe to say that the performance loss from resizing is negligible.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1090463,
          "author_name": "crained",
          "author_url": "",
          "post_date": "11/25/2020 10:54:13",
          "content": "<p>Thank you. I appreciate the response!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1089597": "Love all the great notebooks using Efficientnet but I had a question. I noticed most crop or resize these images to something like 300x300 or 512x512 when the images are 800x600. \n\nThoughts on losing data when doing this? Or just using the same aspect ratio to scale these down? I'm not as familiar with Efficientnet so wanted to get some thoughts. Thanks!\nCharlie",
    "1089847": "The largest efficientnet (B7) was trained on 512x512 images, which is why a lot of people are resizing to that shape. Moreover, it was trained on resized images from ImageNet and was able to achieve 85%+ accuracy, so it's safe to say that the performance loss from resizing is negligible.",
    "1090463": "Thank you. I appreciate the response!"
  },
  "source": "meta"
}