{
  "id": 486197,
  "title": "Data augmentation that can enhance performance in my code",
  "url": "/competitions/hms-harmful-brain-activity-classification/discussion/486197",
  "author_name": "",
  "post_date": "2024-03-23T23:45:49.001146300Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>The data augmentation, as I understand it, for example in the case of images, involves operations such as cropping, masking, etc., to generate more images. This way, during modeling, there can be a greater variety of samples.</p>\n<p>However, I have noticed that many open-source codes transform the data and model with the transformed data. This seems to me like regularization, avoiding overfitting, and possibly improving the model's generalization ability.</p>\n<p><strong>I tried merging the transformed samples with the original ones and then modeling, and there was indeed an improvement in performance, roughly around 1%.</strong></p>\n<p>However, when I assigned relatively small weights to the transformed samples, the model's performance didn't change significantly.</p>\n<p>I hope this technique is helpful for you.</p>",
  "messages": [
    {
      "id": "2713045",
      "postDate": "03/23/2024 23:45:49",
      "content": "<p>The data augmentation, as I understand it, for example in the case of images, involves operations such as cropping, masking, etc., to generate more images. This way, during modeling, there can be a greater variety of samples.</p>\n<p>However, I have noticed that many open-source codes transform the data and model with the transformed data. This seems to me like regularization, avoiding overfitting, and possibly improving the model's generalization ability.</p>\n<p><strong>I tried merging the transformed samples with the original ones and then modeling, and there was indeed an improvement in performance, roughly around 1%.</strong></p>\n<p>However, when I assigned relatively small weights to the transformed samples, the model's performance didn't change significantly.</p>\n<p>I hope this technique is helpful for you.</p>",
      "rawMarkdown": "The data augmentation, as I understand it, for example in the case of images, involves operations such as cropping, masking, etc., to generate more images. This way, during modeling, there can be a greater variety of samples.\n\nHowever, I have noticed that many open-source codes transform the data and model with the transformed data. This seems to me like regularization, avoiding overfitting, and possibly improving the model's generalization ability.\n\n**I tried merging the transformed samples with the original ones and then modeling, and there was indeed an improvement in performance, roughly around 1%.**\n\nHowever, when I assigned relatively small weights to the transformed samples, the model's performance didn't change significantly.\n\nI hope this technique is helpful for you.",
      "votes": null
    },
    {
      "id": "2718839",
      "postDate": "03/27/2024 09:16:57",
      "content": "<p>thanks for sharing, it inspires me a new way to use data augmentation teq😃</p>",
      "rawMarkdown": "thanks for sharing, it inspires me a new way to use data augmentation teq😃",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2718839,
      "author_name": "roger92",
      "author_url": "",
      "post_date": "03/27/2024 09:16:57",
      "content": "<p>thanks for sharing, it inspires me a new way to use data augmentation teq😃</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2713045": "The data augmentation, as I understand it, for example in the case of images, involves operations such as cropping, masking, etc., to generate more images. This way, during modeling, there can be a greater variety of samples.\n\nHowever, I have noticed that many open-source codes transform the data and model with the transformed data. This seems to me like regularization, avoiding overfitting, and possibly improving the model's generalization ability.\n\n**I tried merging the transformed samples with the original ones and then modeling, and there was indeed an improvement in performance, roughly around 1%.**\n\nHowever, when I assigned relatively small weights to the transformed samples, the model's performance didn't change significantly.\n\nI hope this technique is helpful for you.",
    "2718839": "thanks for sharing, it inspires me a new way to use data augmentation teq😃"
  },
  "source": "meta"
}