{
  "id": 473770,
  "title": "Model ensemble or feature ensemble question",
  "url": "/competitions/hms-harmful-brain-activity-classification/discussion/473770",
  "author_name": "",
  "post_date": "2024-02-06T03:07:19.626649600Z",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I am working on some basic models like Resnet101 and Xception and i am wondering that is it possible to combine the feature maps from pretrained models an combine them as input images. <br>\nIs it possible to improve the result? Have anyone try it?<br>\nThanks! :)</p>",
  "messages": [
    {
      "id": "2637960",
      "postDate": "02/06/2024 03:07:19",
      "content": "<p>I am working on some basic models like Resnet101 and Xception and i am wondering that is it possible to combine the feature maps from pretrained models an combine them as input images. <br>\nIs it possible to improve the result? Have anyone try it?<br>\nThanks! :)</p>",
      "rawMarkdown": "I am working on some basic models like Resnet101 and Xception and i am wondering that is it possible to combine the feature maps from pretrained models an combine them as input images. \nIs it possible to improve the result? Have anyone try it?\nThanks! :)",
      "votes": null
    },
    {
      "id": "2638856",
      "postDate": "02/06/2024 14:20:33",
      "content": "<p>In my experiment, the LB of the pretrained model's architecture is boosted by blending.<br>\nthe result of blending tf_efficientnetv2_s/m/l is that LB is boosted by blending compared to submitting in single.</p>\n<ul>\n<li>see: Public My Blend Kernel</li>\n<li><a href=\"https://www.kaggle.com/code/hideyukizushi/hms-only-spectrograms-efficientnetv2-s-m-l-lb-42\" target=\"_blank\">https://www.kaggle.com/code/hideyukizushi/hms-only-spectrograms-efficientnetv2-s-m-l-lb-42</a></li>\n</ul>",
      "rawMarkdown": "In my experiment, the LB of the pretrained model's architecture is boosted by blending.\nthe result of blending tf_efficientnetv2_s/m/l is that LB is boosted by blending compared to submitting in single.\n* see: Public My Blend Kernel\n* https://www.kaggle.com/code/hideyukizushi/hms-only-spectrograms-efficientnetv2-s-m-l-lb-42",
      "votes": null
    },
    {
      "id": "2645735",
      "postDate": "02/10/2024 12:39:27",
      "content": "<p>Thanks!!!@yukiZ, I will try it later.</p>",
      "rawMarkdown": "Thanks!!!@yukiZ, I will try it later.",
      "votes": null
    },
    {
      "id": "2656172",
      "postDate": "02/17/2024 14:12:27",
      "content": "<p>do you mean 'blending blending' or 'simple ensembling by average/max etc.' ?</p>",
      "rawMarkdown": "do you mean 'blending blending' or 'simple ensembling by average/max etc.' ?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2638856,
      "author_name": "hideyukizushi",
      "author_url": "",
      "post_date": "02/06/2024 14:20:33",
      "content": "<p>In my experiment, the LB of the pretrained model's architecture is boosted by blending.<br>\nthe result of blending tf_efficientnetv2_s/m/l is that LB is boosted by blending compared to submitting in single.</p>\n<ul>\n<li>see: Public My Blend Kernel</li>\n<li><a href=\"https://www.kaggle.com/code/hideyukizushi/hms-only-spectrograms-efficientnetv2-s-m-l-lb-42\" target=\"_blank\">https://www.kaggle.com/code/hideyukizushi/hms-only-spectrograms-efficientnetv2-s-m-l-lb-42</a></li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 2645735,
          "author_name": "rexdeng",
          "author_url": "",
          "post_date": "02/10/2024 12:39:27",
          "content": "<p>Thanks!!!@yukiZ, I will try it later.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 2656172,
          "author_name": "abdulkadirguner",
          "author_url": "",
          "post_date": "02/17/2024 14:12:27",
          "content": "<p>do you mean 'blending blending' or 'simple ensembling by average/max etc.' ?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2637960": "I am working on some basic models like Resnet101 and Xception and i am wondering that is it possible to combine the feature maps from pretrained models an combine them as input images. \nIs it possible to improve the result? Have anyone try it?\nThanks! :)",
    "2638856": "In my experiment, the LB of the pretrained model's architecture is boosted by blending.\nthe result of blending tf_efficientnetv2_s/m/l is that LB is boosted by blending compared to submitting in single.\n* see: Public My Blend Kernel\n* https://www.kaggle.com/code/hideyukizushi/hms-only-spectrograms-efficientnetv2-s-m-l-lb-42",
    "2645735": "Thanks!!!@yukiZ, I will try it later.",
    "2656172": "do you mean 'blending blending' or 'simple ensembling by average/max etc.' ?"
  },
  "source": "meta"
}