{
  "id": 590517,
  "title": " Seeking Advice on Improving IMU-only Model (Current PB: 0.79)",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/590517",
  "author_name": "shixingkong",
  "post_date": "2025-07-21T12:29:43.194000",
  "votes": 0,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi Kagglers!</p>\n<p>I’m working on an IMU-only approach and would love insights on further optimizations. Here’s my current setup:</p>\n<p>Current Pipeline:<br>\n✅ Model: Hybrid CNN + LSTM + Attention (similar to this notebook<a href=\"url\" target=\"_blank\">https://www.kaggle.com/code/jiazhuang/cmi-imu-only-lstm</a>)<br>\n✅ Features: Quaternion + acceleration (basic derivations like velocity/distance, inspired by this work<a href=\"url\" target=\"_blank\">https://www.kaggle.com/code/nksusth/lb-0-78-quaternions-tf-bilstm-gru-attention</a>)<br>\n✅ Augmentations: Time-stretch/shift, noise, scaling, IMU rotation, masking, frequency filtering<br>\n✅ Regularization: MixUp for synthetic sample generation</p>\n<p>To IMU-Only Teams: What pushed your score beyond 0.8? Any small wins worth sharing?<br>\nThanks in advance! Happy to exchange ideas/code. 🚀</p>",
  "messages": [
    {
      "id": 3251700,
      "postDate": "2025-07-21T12:29:43.193Z",
      "content": "<p>Hi Kagglers!</p>\n<p>I’m working on an IMU-only approach and would love insights on further optimizations. Here’s my current setup:</p>\n<p>Current Pipeline:<br>\n✅ Model: Hybrid CNN + LSTM + Attention (similar to this notebook<a href=\"url\" target=\"_blank\">https://www.kaggle.com/code/jiazhuang/cmi-imu-only-lstm</a>)<br>\n✅ Features: Quaternion + acceleration (basic derivations like velocity/distance, inspired by this work<a href=\"url\" target=\"_blank\">https://www.kaggle.com/code/nksusth/lb-0-78-quaternions-tf-bilstm-gru-attention</a>)<br>\n✅ Augmentations: Time-stretch/shift, noise, scaling, IMU rotation, masking, frequency filtering<br>\n✅ Regularization: MixUp for synthetic sample generation</p>\n<p>To IMU-Only Teams: What pushed your score beyond 0.8? Any small wins worth sharing?<br>\nThanks in advance! Happy to exchange ideas/code. 🚀</p>",
      "rawMarkdown": "Hi Kagglers!\n\nI’m working on an IMU-only approach and would love insights on further optimizations. Here’s my current setup:\n\nCurrent Pipeline:\n✅ Model: Hybrid CNN + LSTM + Attention (similar to this notebook[https://www.kaggle.com/code/jiazhuang/cmi-imu-only-lstm](url))\n✅ Features: Quaternion + acceleration (basic derivations like velocity/distance, inspired by this work[https://www.kaggle.com/code/nksusth/lb-0-78-quaternions-tf-bilstm-gru-attention](url))\n✅ Augmentations: Time-stretch/shift, noise, scaling, IMU rotation, masking, frequency filtering\n✅ Regularization: MixUp for synthetic sample generation\n\nTo IMU-Only Teams: What pushed your score beyond 0.8? Any small wins worth sharing?\nThanks in advance! Happy to exchange ideas/code. 🚀"
    },
    {
      "id": 3251731,
      "postDate": "2025-07-21T13:28:29.753Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true,
      "replies": [
        {
          "id": 3252100,
          "postDate": "2025-07-22T03:27:52.550Z",
          "content": "<p>It remains unclear whether all data augmentation methods are effective. Recently, I have focused more on feature engineering, as high-quality features have been shown to significantly improve model performance. If time allows, I intend to conduct ablation studies to assess the effectiveness of various data augmentation techniques. </p>",
          "rawMarkdown": "It remains unclear whether all data augmentation methods are effective. Recently, I have focused more on feature engineering, as high-quality features have been shown to significantly improve model performance. If time allows, I intend to conduct ablation studies to assess the effectiveness of various data augmentation techniques. "
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 3251731,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-07-21T13:28:29.753000",
      "content": "",
      "votes": 1,
      "replies": [
        {
          "id": 3252100,
          "author_name": "shixingkong",
          "author_url": "",
          "post_date": "2025-07-22T03:27:52.550000",
          "content": "<p>It remains unclear whether all data augmentation methods are effective. Recently, I have focused more on feature engineering, as high-quality features have been shown to significantly improve model performance. If time allows, I intend to conduct ablation studies to assess the effectiveness of various data augmentation techniques. </p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3251700": "Hi Kagglers!\n\nI’m working on an IMU-only approach and would love insights on further optimizations. Here’s my current setup:\n\nCurrent Pipeline:\n✅ Model: Hybrid CNN + LSTM + Attention (similar to this notebook[https://www.kaggle.com/code/jiazhuang/cmi-imu-only-lstm](url))\n✅ Features: Quaternion + acceleration (basic derivations like velocity/distance, inspired by this work[https://www.kaggle.com/code/nksusth/lb-0-78-quaternions-tf-bilstm-gru-attention](url))\n✅ Augmentations: Time-stretch/shift, noise, scaling, IMU rotation, masking, frequency filtering\n✅ Regularization: MixUp for synthetic sample generation\n\nTo IMU-Only Teams: What pushed your score beyond 0.8? Any small wins worth sharing?\nThanks in advance! Happy to exchange ideas/code. 🚀",
    "3251731": ""
  }
}