{
  "id": 114773,
  "title": "why the benchmark is surpassed within 4 hrs after launch?",
  "url": "/competitions/tensorflow2-question-answering/discussion/114773",
  "author_name": "hengck23",
  "post_date": "2019-10-29T06:47:01.646000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>within such a short time, it is not possible to do any training.\nso i wonder why the benchmark can be surpassed?</p>\n\n<p>possible answers:\n- there exists better open source, e.g. from pytorch, CNTK, TF, other github repo\n- there are some simple post processing\n- the given TF2.0/bert baseline is not tuned or not using most recent models</p>",
  "messages": [
    {
      "id": 660500,
      "postDate": "2019-10-29T08:25:47.497Z",
      "content": "<p><a href=\"https://www.kaggle.com/philculliton/using-tensorflow-2-0-w-bert-on-nq\">The baseline code</a> probably produces non-deterministic results, because I only used that.</p>",
      "rawMarkdown": "[The baseline code](https://www.kaggle.com/philculliton/using-tensorflow-2-0-w-bert-on-nq) probably produces non-deterministic results, because I only used that.",
      "votes": 4,
      "replies": [
        {
          "id": 660659,
          "postDate": "2019-10-29T12:43:24.633Z",
          "content": "<p>I agree with you on this point, I used the baseline code too and I surpassed the benchmark with it.</p>",
          "rawMarkdown": "I agree with you on this point, I used the baseline code too and I surpassed the benchmark with it."
        }
      ]
    },
    {
      "id": 660452,
      "postDate": "2019-10-29T06:47:01.647Z",
      "content": "<p>within such a short time, it is not possible to do any training.\nso i wonder why the benchmark can be surpassed?</p>\n\n<p>possible answers:\n- there exists better open source, e.g. from pytorch, CNTK, TF, other github repo\n- there are some simple post processing\n- the given TF2.0/bert baseline is not tuned or not using most recent models</p>",
      "rawMarkdown": "within such a short time, it is not possible to do any training.\nso i wonder why the benchmark can be surpassed?\n\npossible answers:\n- there exists better open source, e.g. from pytorch, CNTK, TF, other github repo\n- there are some simple post processing\n- the given TF2.0/bert baseline is not tuned or not using most recent models\n\n",
      "votes": 2
    },
    {
      "id": 660525,
      "postDate": "2019-10-29T09:06:48.197Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 660500,
      "author_name": "Tony Y.",
      "author_url": "",
      "post_date": "2019-10-29T08:25:47.497000",
      "content": "<p><a href=\"https://www.kaggle.com/philculliton/using-tensorflow-2-0-w-bert-on-nq\">The baseline code</a> probably produces non-deterministic results, because I only used that.</p>",
      "votes": 4,
      "replies": [
        {
          "id": 660659,
          "author_name": "H4k333m",
          "author_url": "",
          "post_date": "2019-10-29T12:43:24.633000",
          "content": "<p>I agree with you on this point, I used the baseline code too and I surpassed the benchmark with it.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 660525,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-10-29T09:06:48.197000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "660500": "[The baseline code](https://www.kaggle.com/philculliton/using-tensorflow-2-0-w-bert-on-nq) probably produces non-deterministic results, because I only used that.",
    "660452": "within such a short time, it is not possible to do any training.\nso i wonder why the benchmark can be surpassed?\n\npossible answers:\n- there exists better open source, e.g. from pytorch, CNTK, TF, other github repo\n- there are some simple post processing\n- the given TF2.0/bert baseline is not tuned or not using most recent models\n\n",
    "660525": ""
  }
}