{
  "id": 238828,
  "title": "16bit or 8bit images? Does it make a difference?",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/238828",
  "author_name": "CroDoc",
  "post_date": "2021-05-13T15:09:43.636000",
  "votes": 5,
  "comment_count": 6,
  "views": 0,
  "content": "<p>This is something that bugged me throughout the contest … but never tested it? Anyone compared results using 16bit vs 8bit images? Does it make a difference?</p>\n<p>What did you use? We used 16bit … which consumed a lot of memory …</p>",
  "messages": [
    {
      "id": 1305997,
      "postDate": "2021-05-13T15:09:43.637Z",
      "content": "<p>This is something that bugged me throughout the contest … but never tested it? Anyone compared results using 16bit vs 8bit images? Does it make a difference?</p>\n<p>What did you use? We used 16bit … which consumed a lot of memory …</p>",
      "rawMarkdown": "This is something that bugged me throughout the contest ... but never tested it? Anyone compared results using 16bit vs 8bit images? Does it make a difference?\n\nWhat did you use? We used 16bit ... which consumed a lot of memory ...",
      "votes": 5
    },
    {
      "id": 1307031,
      "postDate": "2021-05-14T08:02:01.700Z",
      "content": "<p>Does this mean we were the only one using 16bit? 🤦‍♂️</p>",
      "rawMarkdown": "Does this mean we were the only one using 16bit? 🤦‍♂️",
      "votes": 1
    },
    {
      "id": 1307026,
      "postDate": "2021-05-14T07:58:23.937Z",
      "content": "<p>8 bits to compress the size and speed up iteration. For the same reason, until the very end of the competition, my working model was a single plain-vanilla resnet 18.</p>",
      "rawMarkdown": "8 bits to compress the size and speed up iteration. For the same reason, until the very end of the competition, my working model was a single plain-vanilla resnet 18.",
      "votes": 2
    },
    {
      "id": 1306022,
      "postDate": "2021-05-13T15:18:49.060Z",
      "content": "<p>8-bit, never really considered using 16-bit. I assumed that since we're scaling the image resolutions down anyway, the value of the extra precision (if any) would probably already be lost.</p>",
      "rawMarkdown": "8-bit, never really considered using 16-bit. I assumed that since we're scaling the image resolutions down anyway, the value of the extra precision (if any) would probably already be lost.",
      "votes": 2
    },
    {
      "id": 1307122,
      "postDate": "2021-05-14T09:07:37.897Z",
      "content": "<p>Adding 2 image level models trained on 16bit images into my ensemble gives me a minor boost on public LB. However, I cannot tell whether the improvement is contributed by 16bit images or just by the extra 2 models. 😀</p>",
      "rawMarkdown": "Adding 2 image level models trained on 16bit images into my ensemble gives me a minor boost on public LB. However, I cannot tell whether the improvement is contributed by 16bit images or just by the extra 2 models. 😀"
    },
    {
      "id": 1306156,
      "postDate": "2021-05-13T16:29:16.297Z",
      "content": "<p>16 bit loading and processing will cause CPU bottleneck in my setup. </p>",
      "rawMarkdown": "16 bit loading and processing will cause CPU bottleneck in my setup. "
    },
    {
      "id": 1306125,
      "postDate": "2021-05-13T16:09:35.623Z",
      "content": "<p>8 bit always </p>",
      "rawMarkdown": "8 bit always "
    }
  ],
  "comments": [
    {
      "id": 1307031,
      "author_name": "CroDoc",
      "author_url": "",
      "post_date": "2021-05-14T08:02:01.700000",
      "content": "<p>Does this mean we were the only one using 16bit? 🤦‍♂️</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1307026,
      "author_name": "narsil (jobs-in-data.com)",
      "author_url": "",
      "post_date": "2021-05-14T07:58:23.937000",
      "content": "<p>8 bits to compress the size and speed up iteration. For the same reason, until the very end of the competition, my working model was a single plain-vanilla resnet 18.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1306022,
      "author_name": "Andrew Tratz",
      "author_url": "",
      "post_date": "2021-05-13T15:18:49.060000",
      "content": "<p>8-bit, never really considered using 16-bit. I assumed that since we're scaling the image resolutions down anyway, the value of the extra precision (if any) would probably already be lost.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1307122,
      "author_name": "Hung Quoc To",
      "author_url": "",
      "post_date": "2021-05-14T09:07:37.897000",
      "content": "<p>Adding 2 image level models trained on 16bit images into my ensemble gives me a minor boost on public LB. However, I cannot tell whether the improvement is contributed by 16bit images or just by the extra 2 models. 😀</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1306156,
      "author_name": "Guanshuo Xu",
      "author_url": "",
      "post_date": "2021-05-13T16:29:16.297000",
      "content": "<p>16 bit loading and processing will cause CPU bottleneck in my setup. </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1306125,
      "author_name": "yuvaramsingh",
      "author_url": "",
      "post_date": "2021-05-13T16:09:35.623000",
      "content": "<p>8 bit always </p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1305997": "This is something that bugged me throughout the contest ... but never tested it? Anyone compared results using 16bit vs 8bit images? Does it make a difference?\n\nWhat did you use? We used 16bit ... which consumed a lot of memory ...",
    "1307031": "Does this mean we were the only one using 16bit? 🤦‍♂️",
    "1307026": "8 bits to compress the size and speed up iteration. For the same reason, until the very end of the competition, my working model was a single plain-vanilla resnet 18.",
    "1306022": "8-bit, never really considered using 16-bit. I assumed that since we're scaling the image resolutions down anyway, the value of the extra precision (if any) would probably already be lost.",
    "1307122": "Adding 2 image level models trained on 16bit images into my ensemble gives me a minor boost on public LB. However, I cannot tell whether the improvement is contributed by 16bit images or just by the extra 2 models. 😀",
    "1306156": "16 bit loading and processing will cause CPU bottleneck in my setup. ",
    "1306125": "8 bit always "
  }
}