{
  "id": 464500,
  "title": "Resizing the inference masks back to original scale without LB score loss?",
  "url": "/competitions/blood-vessel-segmentation/discussion/464500",
  "author_name": "",
  "post_date": "2023-12-30T18:37:58.538342100Z",
  "votes": 3,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi, I've been reading that a lot of people have been scaling down the test images pre inference down to 800x800, making the inference, resizing the mask back to whatever the size of the image was, then submitting.I understand that for certain images your forced to resize them because their scales aren't divisible by 16 causing tensor errors, but how do you avoid score loss? when you resize your masks doesnt that distort them?</p>",
  "messages": [
    {
      "id": "2580384",
      "postDate": "12/30/2023 18:37:58",
      "content": "<p>Hi, I've been reading that a lot of people have been scaling down the test images pre inference down to 800x800, making the inference, resizing the mask back to whatever the size of the image was, then submitting.I understand that for certain images your forced to resize them because their scales aren't divisible by 16 causing tensor errors, but how do you avoid score loss? when you resize your masks doesnt that distort them?</p>",
      "rawMarkdown": "Hi, I've been reading that a lot of people have been scaling down the test images pre inference down to 800x800, making the inference, resizing the mask back to whatever the size of the image was, then submitting.I understand that for certain images your forced to resize them because their scales aren't divisible by 16 causing tensor errors, but how do you avoid score loss? when you resize your masks doesnt that distort them?",
      "votes": null
    },
    {
      "id": "2580401",
      "postDate": "12/30/2023 19:21:18",
      "content": "<p>Of course. The only situation I can imagine for downscale previous inference is to avoid memory limitations. If you want inference over the whole kidney at once for example. It will be then a matter of what affects more if the error when upscaling or the exactitude of avoiding patch ensembles. </p>",
      "rawMarkdown": "Of course. The only situation I can imagine for downscale previous inference is to avoid memory limitations. If you want inference over the whole kidney at once for example. It will be then a matter of what affects more if the error when upscaling or the exactitude of avoiding patch ensembles.",
      "votes": null
    },
    {
      "id": "2580477",
      "postDate": "12/30/2023 20:56:40",
      "content": "<p>I see. during training do you not resize your images though?</p>",
      "rawMarkdown": "I see. during training do you not resize your images though?",
      "votes": null
    },
    {
      "id": "2580478",
      "postDate": "12/30/2023 21:01:01",
      "content": "<p>For now I train on crops big enough to direct inference on the full images.</p>",
      "rawMarkdown": "For now I train on crops big enough to direct inference on the full images.",
      "votes": null
    },
    {
      "id": "2580479",
      "postDate": "12/30/2023 21:07:17",
      "content": "<p>I assume you do a random crop with 100%. Would you mind telling me the dimensions you used for your crops?</p>",
      "rawMarkdown": "I assume you do a random crop with 100%. Would you mind telling me the dimensions you used for your crops?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2580401,
      "author_name": "sacuscreed",
      "author_url": "",
      "post_date": "12/30/2023 19:21:18",
      "content": "<p>Of course. The only situation I can imagine for downscale previous inference is to avoid memory limitations. If you want inference over the whole kidney at once for example. It will be then a matter of what affects more if the error when upscaling or the exactitude of avoiding patch ensembles. </p>",
      "votes": null,
      "replies": [
        {
          "id": 2580477,
          "author_name": "zuni8789798",
          "author_url": "",
          "post_date": "12/30/2023 20:56:40",
          "content": "<p>I see. during training do you not resize your images though?</p>",
          "votes": null,
          "replies": [
            {
              "id": 2580478,
              "author_name": "sacuscreed",
              "author_url": "",
              "post_date": "12/30/2023 21:01:01",
              "content": "<p>For now I train on crops big enough to direct inference on the full images.</p>",
              "votes": null,
              "replies": [
                {
                  "id": 2580479,
                  "author_name": "zuni8789798",
                  "author_url": "",
                  "post_date": "12/30/2023 21:07:17",
                  "content": "<p>I assume you do a random crop with 100%. Would you mind telling me the dimensions you used for your crops?</p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2580384": "Hi, I've been reading that a lot of people have been scaling down the test images pre inference down to 800x800, making the inference, resizing the mask back to whatever the size of the image was, then submitting.I understand that for certain images your forced to resize them because their scales aren't divisible by 16 causing tensor errors, but how do you avoid score loss? when you resize your masks doesnt that distort them?",
    "2580401": "Of course. The only situation I can imagine for downscale previous inference is to avoid memory limitations. If you want inference over the whole kidney at once for example. It will be then a matter of what affects more if the error when upscaling or the exactitude of avoiding patch ensembles.",
    "2580477": "I see. during training do you not resize your images though?",
    "2580478": "For now I train on crops big enough to direct inference on the full images.",
    "2580479": "I assume you do a random crop with 100%. Would you mind telling me the dimensions you used for your crops?"
  },
  "source": "meta"
}