{
  "id": 302624,
  "title": "Why does a higher resolution get a higher score？？？",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/302624",
  "author_name": "",
  "post_date": "2022-01-23T12:29:00.496197100Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Q1 : When I tested the model in high resolution, there were a lot of \"fake starfish\" on the picture, but why did I get a higher score?</p>\n<p>Q2 : High resolution means more inference time, private data is 4 times as much as public data, so will a submission that uses high resolution time out (notebook&lt;9 hours) ?</p>",
  "messages": [
    {
      "id": "1661372",
      "postDate": "01/23/2022 12:29:00",
      "content": "<p>Q1 : When I tested the model in high resolution, there were a lot of \"fake starfish\" on the picture, but why did I get a higher score?</p>\n<p>Q2 : High resolution means more inference time, private data is 4 times as much as public data, so will a submission that uses high resolution time out (notebook&lt;9 hours) ?</p>",
      "rawMarkdown": "Q1 : When I tested the model in high resolution, there were a lot of \"fake starfish\" on the picture, but why did I get a higher score?\n\nQ2 : High resolution means more inference time, private data is 4 times as much as public data, so will a submission that uses high resolution time out (notebook<9 hours) ?",
      "votes": null
    },
    {
      "id": "1661385",
      "postDate": "01/23/2022 12:37:04",
      "content": "<p>F2 score metric is used in this competition. The F2 metric weights recall more heavily than precision, as in this case it makes sense to tolerate some false positives in order to ensure very few starfish are missed. That's why you observe more FP with improved KPI.</p>\n<p>What about the second question - obviously processing more data requires more time/computation resources. Whether you run out of time really depends on your model. Some models are faster, some will run out of memory…</p>",
      "rawMarkdown": "F2 score metric is used in this competition. The F2 metric weights recall more heavily than precision, as in this case it makes sense to tolerate some false positives in order to ensure very few starfish are missed. That's why you observe more FP with improved KPI.\n\nWhat about the second question - obviously processing more data requires more time/computation resources. Whether you run out of time really depends on your model. Some models are faster, some will run out of memory...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1661385,
      "author_name": "meowmeowmeowmeowmeow",
      "author_url": "",
      "post_date": "01/23/2022 12:37:04",
      "content": "<p>F2 score metric is used in this competition. The F2 metric weights recall more heavily than precision, as in this case it makes sense to tolerate some false positives in order to ensure very few starfish are missed. That's why you observe more FP with improved KPI.</p>\n<p>What about the second question - obviously processing more data requires more time/computation resources. Whether you run out of time really depends on your model. Some models are faster, some will run out of memory…</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1661372": "Q1 : When I tested the model in high resolution, there were a lot of \"fake starfish\" on the picture, but why did I get a higher score?\n\nQ2 : High resolution means more inference time, private data is 4 times as much as public data, so will a submission that uses high resolution time out (notebook<9 hours) ?",
    "1661385": "F2 score metric is used in this competition. The F2 metric weights recall more heavily than precision, as in this case it makes sense to tolerate some false positives in order to ensure very few starfish are missed. That's why you observe more FP with improved KPI.\n\nWhat about the second question - obviously processing more data requires more time/computation resources. Whether you run out of time really depends on your model. Some models are faster, some will run out of memory..."
  },
  "source": "meta"
}