{
  "id": 480935,
  "title": "Model embeddings",
  "url": "/competitions/hms-harmful-brain-activity-classification/discussion/480935",
  "author_name": "stefanoclss",
  "post_date": "2024-03-01T12:48:40.037000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>During this competition, my goal up to now was to take a look on how to effectively combine the outputs of multiple well performing models. I only relied on CV to publish to the leaderboard and I got 0.47 by optimizing the weights between this 10 models and 0.42 almost equal to the best performing models by introducing a meta model that takes in the the predictions for each fold.  Has anyone tried combining models by using the second to last layer (The embedding layer after the final maxpool layer in the effnet starter). I'm eager to test this but GPU-time is not unlimited.</p>",
  "messages": [
    {
      "id": 2676338,
      "postDate": "2024-03-01T12:48:40.037Z",
      "content": "<p>During this competition, my goal up to now was to take a look on how to effectively combine the outputs of multiple well performing models. I only relied on CV to publish to the leaderboard and I got 0.47 by optimizing the weights between this 10 models and 0.42 almost equal to the best performing models by introducing a meta model that takes in the the predictions for each fold.  Has anyone tried combining models by using the second to last layer (The embedding layer after the final maxpool layer in the effnet starter). I'm eager to test this but GPU-time is not unlimited.</p>",
      "rawMarkdown": "During this competition, my goal up to now was to take a look on how to effectively combine the outputs of multiple well performing models. I only relied on CV to publish to the leaderboard and I got 0.47 by optimizing the weights between this 10 models and 0.42 almost equal to the best performing models by introducing a meta model that takes in the the predictions for each fold.  Has anyone tried combining models by using the second to last layer (The embedding layer after the final maxpool layer in the effnet starter). I'm eager to test this but GPU-time is not unlimited."
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2676338": "During this competition, my goal up to now was to take a look on how to effectively combine the outputs of multiple well performing models. I only relied on CV to publish to the leaderboard and I got 0.47 by optimizing the weights between this 10 models and 0.42 almost equal to the best performing models by introducing a meta model that takes in the the predictions for each fold.  Has anyone tried combining models by using the second to last layer (The embedding layer after the final maxpool layer in the effnet starter). I'm eager to test this but GPU-time is not unlimited."
  }
}