{
  "id": 433386,
  "title": "Anyone implemented ensembling CTC models?",
  "url": "/competitions/asl-fingerspelling/discussion/433386",
  "author_name": "Md Boktiar Mahbub Murad",
  "post_date": "2023-08-21T14:26:10.382000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi, I was facing trouble while ensembling CTC models and was wondering if anyone ensembled multiple CTC models. It looks like CTC model isn't as straightforward as ensembling regular models, since different models might detect different length of sequences.</p>",
  "messages": [
    {
      "id": 2401269,
      "postDate": "2023-08-21T14:26:10.383Z",
      "content": "<p>Hi, I was facing trouble while ensembling CTC models and was wondering if anyone ensembled multiple CTC models. It looks like CTC model isn't as straightforward as ensembling regular models, since different models might detect different length of sequences.</p>",
      "rawMarkdown": "Hi, I was facing trouble while ensembling CTC models and was wondering if anyone ensembled multiple CTC models. It looks like CTC model isn't as straightforward as ensembling regular models, since different models might detect different length of sequences.",
      "votes": 1
    },
    {
      "id": 2404098,
      "postDate": "2023-08-23T03:56:50.563Z",
      "content": "<p>there are many ways to ensemble. averaging is only one of the way. <br>\nyou can google for more, e.g.</p>\n<ul>\n<li>build an ensemble mini network to use features from base model, </li>\n<li>re-score the predicted seq from base models and choose the base,</li>\n<li>distill and use base models as teacher, etc …. re-label… force alignement …</li>\n<li>switching/ MoE</li>\n</ul>",
      "rawMarkdown": "there are many ways to ensemble. averaging is only one of the way. \nyou can google for more, e.g.\n- build an ensemble mini network to use features from base model, \n- re-score the predicted seq from base models and choose the base,\n- distill and use base models as teacher, etc .... re-label... force alignement ...\n- switching/ MoE"
    }
  ],
  "comments": [
    {
      "id": 2404098,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2023-08-23T03:56:50.563000",
      "content": "<p>there are many ways to ensemble. averaging is only one of the way. <br>\nyou can google for more, e.g.</p>\n<ul>\n<li>build an ensemble mini network to use features from base model, </li>\n<li>re-score the predicted seq from base models and choose the base,</li>\n<li>distill and use base models as teacher, etc …. re-label… force alignement …</li>\n<li>switching/ MoE</li>\n</ul>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2401269": "Hi, I was facing trouble while ensembling CTC models and was wondering if anyone ensembled multiple CTC models. It looks like CTC model isn't as straightforward as ensembling regular models, since different models might detect different length of sequences.",
    "2404098": "there are many ways to ensemble. averaging is only one of the way. \nyou can google for more, e.g.\n- build an ensemble mini network to use features from base model, \n- re-score the predicted seq from base models and choose the base,\n- distill and use base models as teacher, etc .... re-label... force alignement ...\n- switching/ MoE"
  }
}