{
  "id": 86536,
  "title": "67th solution: kernel only",
  "url": "/competitions/vsb-power-line-fault-detection/writeups/katsunori-nakai-67th-solution-kernel-only",
  "author_name": "",
  "post_date": "2019-03-25T00:29:10.458957800Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I used only kernel in this competition. \nPlease ask me if you are interested in.</p>\n\n<p><a href=\"https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp\">https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp</a></p>",
  "messages": [
    {
      "id": "499561",
      "postDate": "03/25/2019 00:29:10",
      "content": "<p>I used only kernel in this competition. \nPlease ask me if you are interested in.</p>\n\n<p><a href=\"https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp\">https://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp</a></p>",
      "rawMarkdown": "I used only kernel in this competition. \nPlease ask me if you are interested in.\n\nhttps://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp",
      "votes": null
    },
    {
      "id": "499591",
      "postDate": "03/25/2019 01:41:58",
      "content": "<p>Thanks for sharing.  The StatefullMCC layer is great idea !  </p>",
      "rawMarkdown": "Thanks for sharing.  The StatefullMCC layer is great idea !",
      "votes": null
    },
    {
      "id": "499607",
      "postDate": "03/25/2019 01:58:19",
      "content": "<p>\"The standart way keras calculates an epoch's loss/metric is by taking the average value of the loss/metric on batches. While it works well on losses calculated by average (like cross entropy), it creates problems in metrics that must be calculated over the entire dataset (like F1 score or Matthews Correlation).\nTo overcome that problems, we are going to use a statefull metric that is calculated over a whole epoch. A stateful metric on keras is a special layer that allow the running of cumulative operations on each batch. The following implementation requires tensorflow as backend.\"\n(from shared kernel)</p>\n\n<p>Wow. Great Idea! Thanks for sharing it!</p>",
      "rawMarkdown": "\"The standart way keras calculates an epoch's loss/metric is by taking the average value of the loss/metric on batches. While it works well on losses calculated by average (like cross entropy), it creates problems in metrics that must be calculated over the entire dataset (like F1 score or Matthews Correlation).\nTo overcome that problems, we are going to use a statefull metric that is calculated over a whole epoch. A stateful metric on keras is a special layer that allow the running of cumulative operations on each batch. The following implementation requires tensorflow as backend.\"\n(from shared kernel)\n\nWow. Great Idea! Thanks for sharing it!",
      "votes": null
    },
    {
      "id": "499660",
      "postDate": "03/25/2019 04:36:52",
      "content": "<p>Thanks for sharing ! </p>",
      "rawMarkdown": "Thanks for sharing !",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 499591,
      "author_name": "qinhui1999",
      "author_url": "",
      "post_date": "03/25/2019 01:41:58",
      "content": "<p>Thanks for sharing.  The StatefullMCC layer is great idea !  </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 499607,
      "author_name": "mykim89",
      "author_url": "",
      "post_date": "03/25/2019 01:58:19",
      "content": "<p>\"The standart way keras calculates an epoch's loss/metric is by taking the average value of the loss/metric on batches. While it works well on losses calculated by average (like cross entropy), it creates problems in metrics that must be calculated over the entire dataset (like F1 score or Matthews Correlation).\nTo overcome that problems, we are going to use a statefull metric that is calculated over a whole epoch. A stateful metric on keras is a special layer that allow the running of cumulative operations on each batch. The following implementation requires tensorflow as backend.\"\n(from shared kernel)</p>\n\n<p>Wow. Great Idea! Thanks for sharing it!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 499660,
      "author_name": "dhaqui",
      "author_url": "",
      "post_date": "03/25/2019 04:36:52",
      "content": "<p>Thanks for sharing ! </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "499561": "I used only kernel in this competition. \nPlease ask me if you are interested in.\n\nhttps://www.kaggle.com/bigswimatom/5-fold-lstm-attention-stateful-metrics-with-exp",
    "499591": "Thanks for sharing.  The StatefullMCC layer is great idea !",
    "499607": "\"The standart way keras calculates an epoch's loss/metric is by taking the average value of the loss/metric on batches. While it works well on losses calculated by average (like cross entropy), it creates problems in metrics that must be calculated over the entire dataset (like F1 score or Matthews Correlation).\nTo overcome that problems, we are going to use a statefull metric that is calculated over a whole epoch. A stateful metric on keras is a special layer that allow the running of cumulative operations on each batch. The following implementation requires tensorflow as backend.\"\n(from shared kernel)\n\nWow. Great Idea! Thanks for sharing it!",
    "499660": "Thanks for sharing !"
  },
  "source": "meta"
}