{
  "id": 78537,
  "title": "How to increase further accuracy.",
  "url": "/competitions/histopathologic-cancer-detection/discussion/78537",
  "author_name": "",
  "post_date": "2019-01-25T05:51:29.507890200Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I am stuck with accuracy of 87% and it seems that my model is trying to label as positive class even though test image does not have cell in 32*32 central areas. How we can ensure that model only looks into desired area of the image and not whole image? Any idea from the peoples who are in top 50 (atleast) will be appreciated and will help us to learn...</p>",
  "messages": [
    {
      "id": "461056",
      "postDate": "01/25/2019 05:51:29",
      "content": "<p>I am stuck with accuracy of 87% and it seems that my model is trying to label as positive class even though test image does not have cell in 32*32 central areas. How we can ensure that model only looks into desired area of the image and not whole image? Any idea from the peoples who are in top 50 (atleast) will be appreciated and will help us to learn...</p>",
      "rawMarkdown": "I am stuck with accuracy of 87% and it seems that my model is trying to label as positive class even though test image does not have cell in 32*32 central areas. How we can ensure that model only looks into desired area of the image and not whole image? Any idea from the peoples who are in top 50 (atleast) will be appreciated and will help us to learn...",
      "votes": null
    },
    {
      "id": "461143",
      "postDate": "01/25/2019 10:49:50",
      "content": "<p>Try different models and architectures, Increase the image size for your input images, and basically train more epochs. You should get a better score. What library are you using?</p>",
      "rawMarkdown": "Try different models and architectures, Increase the image size for your input images, and basically train more epochs. You should get a better score. What library are you using?",
      "votes": null
    },
    {
      "id": "461546",
      "postDate": "01/26/2019 10:05:41",
      "content": "<p>I am using keras. My concern is how i can ensure my network does not label image as positive class when the cell is not in central are. same here during training also.</p>",
      "rawMarkdown": "I am using keras. My concern is how i can ensure my network does not label image as positive class when the cell is not in central are. same here during training also.",
      "votes": null
    },
    {
      "id": "462233",
      "postDate": "01/27/2019 23:17:02",
      "content": "<p>Try data augmentation. What you can do is to apply random translation (and many other image transformations) to your training data. In that way, when your data have cell center in all locations, then your model will not be trained to think that having cell in the central of the image is of a particular class and vice versa.</p>",
      "rawMarkdown": "Try data augmentation. What you can do is to apply random translation (and many other image transformations) to your training data. In that way, when your data have cell center in all locations, then your model will not be trained to think that having cell in the central of the image is of a particular class and vice versa.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 461143,
      "author_name": "abdishakuur",
      "author_url": "",
      "post_date": "01/25/2019 10:49:50",
      "content": "<p>Try different models and architectures, Increase the image size for your input images, and basically train more epochs. You should get a better score. What library are you using?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 461546,
      "author_name": "anyltcamitk",
      "author_url": "",
      "post_date": "01/26/2019 10:05:41",
      "content": "<p>I am using keras. My concern is how i can ensure my network does not label image as positive class when the cell is not in central are. same here during training also.</p>",
      "votes": null,
      "replies": [
        {
          "id": 462233,
          "author_name": "gxkok21",
          "author_url": "",
          "post_date": "01/27/2019 23:17:02",
          "content": "<p>Try data augmentation. What you can do is to apply random translation (and many other image transformations) to your training data. In that way, when your data have cell center in all locations, then your model will not be trained to think that having cell in the central of the image is of a particular class and vice versa.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "461056": "I am stuck with accuracy of 87% and it seems that my model is trying to label as positive class even though test image does not have cell in 32*32 central areas. How we can ensure that model only looks into desired area of the image and not whole image? Any idea from the peoples who are in top 50 (atleast) will be appreciated and will help us to learn...",
    "461143": "Try different models and architectures, Increase the image size for your input images, and basically train more epochs. You should get a better score. What library are you using?",
    "461546": "I am using keras. My concern is how i can ensure my network does not label image as positive class when the cell is not in central are. same here during training also.",
    "462233": "Try data augmentation. What you can do is to apply random translation (and many other image transformations) to your training data. In that way, when your data have cell center in all locations, then your model will not be trained to think that having cell in the central of the image is of a particular class and vice versa."
  },
  "source": "meta"
}