{
  "id": 53565,
  "title": "How to monitor MAP in matterport's implementation at the end of each epoch ?",
  "url": "/competitions/data-science-bowl-2018/discussion/53565",
  "author_name": "",
  "post_date": "2018-04-01T16:23:58.549839200Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>After prediction in each epoch, a lot of post-processing (e.g NMS ) is done for final prediction of mask  So how can we monitor validation MAP using Keras callbacks after each epoch ?</p>",
  "messages": [
    {
      "id": "307455",
      "postDate": "04/01/2018 16:23:58",
      "content": "<p>After prediction in each epoch, a lot of post-processing (e.g NMS ) is done for final prediction of mask  So how can we monitor validation MAP using Keras callbacks after each epoch ?</p>",
      "rawMarkdown": "After prediction in each epoch, a lot of post-processing (e.g NMS ) is done for final prediction of mask  So how can we monitor validation MAP using Keras callbacks after each epoch ?",
      "votes": null
    },
    {
      "id": "307502",
      "postDate": "04/01/2018 20:52:41",
      "content": "<p>My strategy here is to do it in a separate process so that training isn't slowed. I have a watchdog thread + second GPU waiting for new model weights to appear in a directory:</p>\n\n<ol>\n<li>epoch ends =&gt; model's weights are written to a directory</li>\n<li>map monitor proc picks up weights and computes the eval metric</li>\n<li>track mAP for each epoch:</li>\n</ol>",
      "rawMarkdown": "My strategy here is to do it in a separate process so that training isn't slowed. I have a watchdog thread + second GPU waiting for new model weights to appear in a directory:\n\n 1. epoch ends =&gt; model's weights are written to a directory\n 2. map monitor proc picks up weights and computes the eval metric\n 3. track mAP for each epoch:",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 307502,
      "author_name": "spnrpa",
      "author_url": "",
      "post_date": "04/01/2018 20:52:41",
      "content": "<p>My strategy here is to do it in a separate process so that training isn't slowed. I have a watchdog thread + second GPU waiting for new model weights to appear in a directory:</p>\n\n<ol>\n<li>epoch ends =&gt; model's weights are written to a directory</li>\n<li>map monitor proc picks up weights and computes the eval metric</li>\n<li>track mAP for each epoch:</li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "307455": "After prediction in each epoch, a lot of post-processing (e.g NMS ) is done for final prediction of mask  So how can we monitor validation MAP using Keras callbacks after each epoch ?",
    "307502": "My strategy here is to do it in a separate process so that training isn't slowed. I have a watchdog thread + second GPU waiting for new model weights to appear in a directory:\n\n 1. epoch ends =&gt; model's weights are written to a directory\n 2. map monitor proc picks up weights and computes the eval metric\n 3. track mAP for each epoch:"
  },
  "source": "meta"
}