{
  "id": 301664,
  "title": "Trade off inference time",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/301664",
  "author_name": "",
  "post_date": "2022-01-18T21:45:11.575978800Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>The high score kernels is helped by high resolution images. However, due to the large image size, it takes a very long inference time. (4 to 5 hours) In other words, when the image size becomes huge, such as 9000, the maximum number of models that can be ensemble is two, which is disadvantageous in terms of model diversity.</p>\n<p>Currently, I have multiple single models that can get a score of around 0.55 with image size = 1280. (without post-processing) The inference time for one of these models is about 30 minutes. This means that you can ensemble more than 8 models in the time you are running a high resolution single model.</p>\n<p>Is it possible to create a model that fits Private LB with a small number of models?</p>",
  "messages": [
    {
      "id": "1654853",
      "postDate": "01/18/2022 21:45:11",
      "content": "<p>The high score kernels is helped by high resolution images. However, due to the large image size, it takes a very long inference time. (4 to 5 hours) In other words, when the image size becomes huge, such as 9000, the maximum number of models that can be ensemble is two, which is disadvantageous in terms of model diversity.</p>\n<p>Currently, I have multiple single models that can get a score of around 0.55 with image size = 1280. (without post-processing) The inference time for one of these models is about 30 minutes. This means that you can ensemble more than 8 models in the time you are running a high resolution single model.</p>\n<p>Is it possible to create a model that fits Private LB with a small number of models?</p>",
      "rawMarkdown": "The high score kernels is helped by high resolution images. However, due to the large image size, it takes a very long inference time. (4 to 5 hours) In other words, when the image size becomes huge, such as 9000, the maximum number of models that can be ensemble is two, which is disadvantageous in terms of model diversity.\n\nCurrently, I have multiple single models that can get a score of around 0.55 with image size = 1280. (without post-processing) The inference time for one of these models is about 30 minutes. This means that you can ensemble more than 8 models in the time you are running a high resolution single model.\n\nIs it possible to create a model that fits Private LB with a small number of models?",
      "votes": null
    },
    {
      "id": "1659897",
      "postDate": "01/22/2022 07:23:50",
      "content": "<p>It really depends on your model differences in principle. Based on previous competitions, the model should be derived from different approaches (YOLO, R-CNN, etc.) to show huge score performance.</p>",
      "rawMarkdown": "It really depends on your model differences in principle. Based on previous competitions, the model should be derived from different approaches (YOLO, R-CNN, etc.) to show huge score performance.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1659897,
      "author_name": "hsadeghian",
      "author_url": "",
      "post_date": "01/22/2022 07:23:50",
      "content": "<p>It really depends on your model differences in principle. Based on previous competitions, the model should be derived from different approaches (YOLO, R-CNN, etc.) to show huge score performance.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1654853": "The high score kernels is helped by high resolution images. However, due to the large image size, it takes a very long inference time. (4 to 5 hours) In other words, when the image size becomes huge, such as 9000, the maximum number of models that can be ensemble is two, which is disadvantageous in terms of model diversity.\n\nCurrently, I have multiple single models that can get a score of around 0.55 with image size = 1280. (without post-processing) The inference time for one of these models is about 30 minutes. This means that you can ensemble more than 8 models in the time you are running a high resolution single model.\n\nIs it possible to create a model that fits Private LB with a small number of models?",
    "1659897": "It really depends on your model differences in principle. Based on previous competitions, the model should be derived from different approaches (YOLO, R-CNN, etc.) to show huge score performance."
  },
  "source": "meta"
}