{
  "id": 241277,
  "title": "Optimal Image Size",
  "url": "/competitions/bms-molecular-translation/discussion/241277",
  "author_name": "",
  "post_date": "2021-05-23T22:05:29.272196100Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>What Image size are you using for your Models?</p>",
  "messages": [
    {
      "id": "1320228",
      "postDate": "05/23/2021 22:05:29",
      "content": "<p>What Image size are you using for your Models?</p>",
      "rawMarkdown": "What Image size are you using for your Models?",
      "votes": null
    },
    {
      "id": "1320930",
      "postDate": "05/24/2021 12:27:44",
      "content": "<p>I am using 512 * 256 images. The cropped images have an aspect ratio ~2 and the chosen size should have a reasonable training time and fit in memory with large encoder models.</p>",
      "rawMarkdown": "I am using 512 * 256 images. The cropped images have an aspect ratio ~2 and the chosen size should have a reasonable training time and fit in memory with large encoder models.",
      "votes": null
    },
    {
      "id": "1323416",
      "postDate": "05/26/2021 08:05:21",
      "content": "<p>384x384 has proven optimal for me. </p>",
      "rawMarkdown": "384x384 has proven optimal for me.",
      "votes": null
    },
    {
      "id": "1323444",
      "postDate": "05/26/2021 08:27:13",
      "content": "<p>Did you try using the original aspect ratio of 2? Intuitively you would expect this ratio to give the best performance, as a ratio 1 of would result in lots of padding.</p>",
      "rawMarkdown": "Did you try using the original aspect ratio of 2? Intuitively you would expect this ratio to give the best performance, as a ratio 1 of would result in lots of padding.",
      "votes": null
    },
    {
      "id": "1323512",
      "postDate": "05/26/2021 09:31:08",
      "content": "<p>Your intuition makes sense to me too, but in my experiments using a different aspect ratio did not achieve a better CV or LB score. But I think there could be a lot more work done on this, perhaps my application of rotation was insufficient! </p>",
      "rawMarkdown": "Your intuition makes sense to me too, but in my experiments using a different aspect ratio did not achieve a better CV or LB score. But I think there could be a lot more work done on this, perhaps my application of rotation was insufficient!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1320930,
      "author_name": "markwijkhuizen",
      "author_url": "",
      "post_date": "05/24/2021 12:27:44",
      "content": "<p>I am using 512 * 256 images. The cropped images have an aspect ratio ~2 and the chosen size should have a reasonable training time and fit in memory with large encoder models.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1323416,
      "author_name": "talktocharles",
      "author_url": "",
      "post_date": "05/26/2021 08:05:21",
      "content": "<p>384x384 has proven optimal for me. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1323444,
          "author_name": "markwijkhuizen",
          "author_url": "",
          "post_date": "05/26/2021 08:27:13",
          "content": "<p>Did you try using the original aspect ratio of 2? Intuitively you would expect this ratio to give the best performance, as a ratio 1 of would result in lots of padding.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1323512,
          "author_name": "talktocharles",
          "author_url": "",
          "post_date": "05/26/2021 09:31:08",
          "content": "<p>Your intuition makes sense to me too, but in my experiments using a different aspect ratio did not achieve a better CV or LB score. But I think there could be a lot more work done on this, perhaps my application of rotation was insufficient! </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1320228": "What Image size are you using for your Models?",
    "1320930": "I am using 512 * 256 images. The cropped images have an aspect ratio ~2 and the chosen size should have a reasonable training time and fit in memory with large encoder models.",
    "1323416": "384x384 has proven optimal for me.",
    "1323444": "Did you try using the original aspect ratio of 2? Intuitively you would expect this ratio to give the best performance, as a ratio 1 of would result in lots of padding.",
    "1323512": "Your intuition makes sense to me too, but in my experiments using a different aspect ratio did not achieve a better CV or LB score. But I think there could be a lot more work done on this, perhaps my application of rotation was insufficient!"
  },
  "source": "meta"
}