{
  "id": 159640,
  "title": "kfold did not improve acc",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/159640",
  "author_name": "",
  "post_date": "2020-06-18T06:32:37.492049300Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I use kfold( k = 3) and sperate train data by patient id. At the same time, i try to keep same number of images in each fold. After that i train 3 models according to imgs in 3 different folds.\nMy network is Resnet50 and output 1 node with sigmoid. </p>\n\n<p>I get mean of outputs from 3 model' prediction. But only got 0.87 in test. That is almost the same with one model(Resnet50). </p>\n\n<p>Can any body help me.</p>",
  "messages": [
    {
      "id": "891385",
      "postDate": "06/18/2020 06:32:37",
      "content": "<p>I use kfold( k = 3) and sperate train data by patient id. At the same time, i try to keep same number of images in each fold. After that i train 3 models according to imgs in 3 different folds.\nMy network is Resnet50 and output 1 node with sigmoid. </p>\n\n<p>I get mean of outputs from 3 model' prediction. But only got 0.87 in test. That is almost the same with one model(Resnet50). </p>\n\n<p>Can any body help me.</p>",
      "rawMarkdown": "I use kfold( k = 3) and sperate train data by patient id. At the same time, i try to keep same number of images in each fold. After that i train 3 models according to imgs in 3 different folds.\nMy network is Resnet50 and output 1 node with sigmoid. \n\nI get mean of outputs from 3 model' prediction. But only got 0.87 in test. That is almost the same with one model(Resnet50). \n\nCan any body help me.",
      "votes": null
    },
    {
      "id": "891403",
      "postDate": "06/18/2020 06:48:14",
      "content": "<p>It's not a guarantee that averaging over the folds will result in a better performance. It will reduce the variance though. The advantage of using KFold is that you have a more reliable score than having  a single train/test split.</p>",
      "rawMarkdown": "It's not a guarantee that averaging over the folds will result in a better performance. It will reduce the variance though. The advantage of using KFold is that you have a more reliable score than having  a single train/test split.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 891403,
      "author_name": "group16",
      "author_url": "",
      "post_date": "06/18/2020 06:48:14",
      "content": "<p>It's not a guarantee that averaging over the folds will result in a better performance. It will reduce the variance though. The advantage of using KFold is that you have a more reliable score than having  a single train/test split.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "891385": "I use kfold( k = 3) and sperate train data by patient id. At the same time, i try to keep same number of images in each fold. After that i train 3 models according to imgs in 3 different folds.\nMy network is Resnet50 and output 1 node with sigmoid. \n\nI get mean of outputs from 3 model' prediction. But only got 0.87 in test. That is almost the same with one model(Resnet50). \n\nCan any body help me.",
    "891403": "It's not a guarantee that averaging over the folds will result in a better performance. It will reduce the variance though. The advantage of using KFold is that you have a more reliable score than having  a single train/test split."
  },
  "source": "meta"
}