{
  "id": 64245,
  "title": "Fine-tuning Faster-RCNN Atrous Version for OID results in trivial score",
  "url": "/competitions/google-ai-open-images-object-detection-track/discussion/64245",
  "author_name": "",
  "post_date": "2018-08-27T08:46:03.537351500Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I built a solution off-the-shelf at first and got 0.230 LB. But when I try to fine-tune it with this year's data, after 10k images the AP score dropped to 0.000. I'm wondering if this is happening to you guys as well since I'm new to this field and have little knowledge of what the training curve should look like. Is there possibly a bug in my code, or are the datasets too different so that I have to re-train everything from scratch?</p>",
  "messages": [
    {
      "id": "376329",
      "postDate": "08/27/2018 08:46:03",
      "content": "<p>I built a solution off-the-shelf at first and got 0.230 LB. But when I try to fine-tune it with this year's data, after 10k images the AP score dropped to 0.000. I'm wondering if this is happening to you guys as well since I'm new to this field and have little knowledge of what the training curve should look like. Is there possibly a bug in my code, or are the datasets too different so that I have to re-train everything from scratch?</p>",
      "rawMarkdown": "I built a solution off-the-shelf at first and got 0.230 LB. But when I try to fine-tune it with this year's data, after 10k images the AP score dropped to 0.000. I'm wondering if this is happening to you guys as well since I'm new to this field and have little knowledge of what the training curve should look like. Is there possibly a bug in my code, or are the datasets too different so that I have to re-train everything from scratch?",
      "votes": null
    },
    {
      "id": "377304",
      "postDate": "08/28/2018 23:36:28",
      "content": "<p>I don't think there's time left to retrain from scratch, so consider other options. </p>",
      "rawMarkdown": "I don't think there's time left to retrain from scratch, so consider other options.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 377304,
      "author_name": "dzspin",
      "author_url": "",
      "post_date": "08/28/2018 23:36:28",
      "content": "<p>I don't think there's time left to retrain from scratch, so consider other options. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "376329": "I built a solution off-the-shelf at first and got 0.230 LB. But when I try to fine-tune it with this year's data, after 10k images the AP score dropped to 0.000. I'm wondering if this is happening to you guys as well since I'm new to this field and have little knowledge of what the training curve should look like. Is there possibly a bug in my code, or are the datasets too different so that I have to re-train everything from scratch?",
    "377304": "I don't think there's time left to retrain from scratch, so consider other options."
  },
  "source": "meta"
}