{
  "id": 76975,
  "title": "ways to fight imbalanced dataset",
  "url": "/competitions/human-protein-atlas-image-classification/discussion/76975",
  "author_name": "",
  "post_date": "2019-01-08T12:21:07.859308100Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I use pytorch\n1.  direct upsample or downsample\n2.  use weights, WeightedRandomSampler\n3.  use weights, BCEWithLogitsLoss(weights)\n4.  focal loss\n5.  off-the-shelf Multilabel Stratification package</p>\n\n<p>I am curious which way is most effective ...</p>",
  "messages": [
    {
      "id": "452243",
      "postDate": "01/08/2019 12:21:07",
      "content": "<p>I use pytorch\n1.  direct upsample or downsample\n2.  use weights, WeightedRandomSampler\n3.  use weights, BCEWithLogitsLoss(weights)\n4.  focal loss\n5.  off-the-shelf Multilabel Stratification package</p>\n\n<p>I am curious which way is most effective ...</p>",
      "rawMarkdown": "I use pytorch\n1.  direct upsample or downsample\n2.  use weights, WeightedRandomSampler\n3.  use weights, BCEWithLogitsLoss(weights)\n4.  focal loss\n5.  off-the-shelf Multilabel Stratification package\n\nI am curious which way is most effective ...",
      "votes": null
    },
    {
      "id": "452316",
      "postDate": "01/08/2019 15:00:16",
      "content": "<p>For me oversampling is the most effective.</p>\n\n<p>I use 2. during training and 5. to make subsets for cross-validation. Number 3. and 4. have no noticeable effect for me. Didn't try 1. but should be pretty similar to 2.</p>",
      "rawMarkdown": "For me oversampling is the most effective.\n\nI use 2. during training and 5. to make subsets for cross-validation. Number 3. and 4. have no noticeable effect for me. Didn't try 1. but should be pretty similar to 2.",
      "votes": null
    },
    {
      "id": "452664",
      "postDate": "01/09/2019 02:32:38",
      "content": "<p>Thanks for answering, currently WeightedRandomSampler boosts my lb by 0.005</p>",
      "rawMarkdown": "Thanks for answering, currently WeightedRandomSampler boosts my lb by 0.005",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 452316,
      "author_name": "dr1t10",
      "author_url": "",
      "post_date": "01/08/2019 15:00:16",
      "content": "<p>For me oversampling is the most effective.</p>\n\n<p>I use 2. during training and 5. to make subsets for cross-validation. Number 3. and 4. have no noticeable effect for me. Didn't try 1. but should be pretty similar to 2.</p>",
      "votes": null,
      "replies": [
        {
          "id": 452664,
          "author_name": "zhangmiao",
          "author_url": "",
          "post_date": "01/09/2019 02:32:38",
          "content": "<p>Thanks for answering, currently WeightedRandomSampler boosts my lb by 0.005</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "452243": "I use pytorch\n1.  direct upsample or downsample\n2.  use weights, WeightedRandomSampler\n3.  use weights, BCEWithLogitsLoss(weights)\n4.  focal loss\n5.  off-the-shelf Multilabel Stratification package\n\nI am curious which way is most effective ...",
    "452316": "For me oversampling is the most effective.\n\nI use 2. during training and 5. to make subsets for cross-validation. Number 3. and 4. have no noticeable effect for me. Didn't try 1. but should be pretty similar to 2.",
    "452664": "Thanks for answering, currently WeightedRandomSampler boosts my lb by 0.005"
  },
  "source": "meta"
}