{
  "id": 127214,
  "title": "what is the best TF2.0 single model?",
  "url": "/competitions/tensorflow2-question-answering/discussion/127214",
  "author_name": "",
  "post_date": "2020-01-23T00:29:25.582561700Z",
  "votes": 6,
  "comment_count": 14,
  "views": 0,
  "content": "<p>our best is 0.64 private lb <a href=\"https://www.kaggle.com/jiweiliu/jb-tf-2-0\">https://www.kaggle.com/jiweiliu/jb-tf-2-0</a> </p>",
  "messages": [
    {
      "id": "726264",
      "postDate": "01/23/2020 00:29:25",
      "content": "<p>our best is 0.64 private lb <a href=\"https://www.kaggle.com/jiweiliu/jb-tf-2-0\">https://www.kaggle.com/jiweiliu/jb-tf-2-0</a> </p>",
      "rawMarkdown": "our best is 0.64 private lb https://www.kaggle.com/jiweiliu/jb-tf-2-0",
      "votes": null
    },
    {
      "id": "726279",
      "postDate": "01/23/2020 00:41:42",
      "content": "<p>We have several single models at 0.69.  But they are pytorch ;)</p>",
      "rawMarkdown": "We have several single models at 0.69.  But they are pytorch ;)",
      "votes": null
    },
    {
      "id": "726284",
      "postDate": "01/23/2020 00:43:17",
      "content": "<p>Have you implemented joint approach or?</p>",
      "rawMarkdown": "Have you implemented joint approach or?",
      "votes": null
    },
    {
      "id": "726286",
      "postDate": "01/23/2020 00:45:16",
      "content": "<p>Variants of it yes.</p>",
      "rawMarkdown": "Variants of it yes.",
      "votes": null
    },
    {
      "id": "726291",
      "postDate": "01/23/2020 00:46:46",
      "content": "<p>single tf1.0 model with public score and private score both 0.70</p>",
      "rawMarkdown": "single tf1.0 model with public score and private score both 0.70",
      "votes": null
    },
    {
      "id": "726293",
      "postDate": "01/23/2020 00:47:57",
      "content": "<p>I have chosen RoBERTa-large, I got 0.67, because of the kernel time limit, I haven't do ensembling.</p>",
      "rawMarkdown": "I have chosen RoBERTa-large, I got 0.67, because of the kernel time limit, I haven't do ensembling.",
      "votes": null
    },
    {
      "id": "726298",
      "postDate": "01/23/2020 00:50:13",
      "content": "<p>What base model have you chosen? I mean BERT or RoBERTa and...</p>",
      "rawMarkdown": "What base model have you chosen? I mean BERT or RoBERTa and...",
      "votes": null
    },
    {
      "id": "726305",
      "postDate": "01/23/2020 00:55:44",
      "content": "<p>I tried bert, xlnet, spanbert and found WWM BERT Large is the best.</p>",
      "rawMarkdown": "I tried bert, xlnet, spanbert and found WWM BERT Large is the best.",
      "votes": null
    },
    {
      "id": "726329",
      "postDate": "01/23/2020 01:12:34",
      "content": "<p>I have a bert-large with private lb 0.65 using TF2.0. </p>",
      "rawMarkdown": "I have a bert-large with private lb 0.65 using TF2.0.",
      "votes": null
    },
    {
      "id": "726358",
      "postDate": "01/23/2020 01:25:46",
      "content": "<p>my best (TF2.0): Bert large (WWM) with last 2 layers combination (but I didn't use this because of public score)</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F193592%2F3ae56be4d3be988891a0a12fd2bb8cda%2F2020-01-23%2010.22.00.png?generation=1579742652969009&amp;alt=media\" alt=\"\"></p>\n\n<p>I'm surprised that some get over 0.67 in single model...!</p>",
      "rawMarkdown": "my best (TF2.0): Bert large (WWM) with last 2 layers combination (but I didn't use this because of public score)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F193592%2F3ae56be4d3be988891a0a12fd2bb8cda%2F2020-01-23%2010.22.00.png?generation=1579742652969009&amp;alt=media)\n\nI'm surprised that some get over 0.67 in single model...!",
      "votes": null
    },
    {
      "id": "726380",
      "postDate": "01/23/2020 01:39:31",
      "content": "<p><a href=\"/cpmpml\">@cpmpml</a> - Congrats :-) . Jus want to know how many single models you have used for ensembling. What was your approach for ensembling?</p>",
      "rawMarkdown": "cpmpml - Congrats :-) . Jus want to know how many single models you have used for ensembling. What was your approach for ensembling?",
      "votes": null
    },
    {
      "id": "727032",
      "postDate": "01/23/2020 11:55:40",
      "content": "<p>0.71 private lb: <a href=\"https://www.kaggle.com/seesee/submit-full\">https://www.kaggle.com/seesee/submit-full</a></p>",
      "rawMarkdown": "0.71 private lb: https://www.kaggle.com/seesee/submit-full",
      "votes": null
    },
    {
      "id": "727035",
      "postDate": "01/23/2020 11:57:59",
      "content": "<p>0.64 private LB. Are you entering your best private LB score even if you didn't select it as your final submission?</p>",
      "rawMarkdown": "0.64 private LB. Are you entering your best private LB score even if you didn't select it as your final submission?",
      "votes": null
    },
    {
      "id": "727049",
      "postDate": "01/23/2020 12:09:37",
      "content": "<p><a href=\"/s4sarath\">@s4sarath</a> you can check in our summary :D</p>",
      "rawMarkdown": "s4sarath you can check in our summary :D",
      "votes": null
    },
    {
      "id": "727104",
      "postDate": "01/23/2020 13:12:58",
      "content": "<p>Thank you :-)</p>",
      "rawMarkdown": "Thank you :-)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 726279,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "01/23/2020 00:41:42",
      "content": "<p>We have several single models at 0.69.  But they are pytorch ;)</p>",
      "votes": null,
      "replies": [
        {
          "id": 726284,
          "author_name": "yaroshevskiy",
          "author_url": "",
          "post_date": "01/23/2020 00:43:17",
          "content": "<p>Have you implemented joint approach or?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 726286,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "01/23/2020 00:45:16",
          "content": "<p>Variants of it yes.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 726380,
          "author_name": "s4sarath",
          "author_url": "",
          "post_date": "01/23/2020 01:39:31",
          "content": "<p><a href=\"/cpmpml\">@cpmpml</a> - Congrats :-) . Jus want to know how many single models you have used for ensembling. What was your approach for ensembling?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 727049,
          "author_name": "christofhenkel",
          "author_url": "",
          "post_date": "01/23/2020 12:09:37",
          "content": "<p><a href=\"/s4sarath\">@s4sarath</a> you can check in our summary :D</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 727104,
          "author_name": "s4sarath",
          "author_url": "",
          "post_date": "01/23/2020 13:12:58",
          "content": "<p>Thank you :-)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 726291,
      "author_name": "tonyxu",
      "author_url": "",
      "post_date": "01/23/2020 00:46:46",
      "content": "<p>single tf1.0 model with public score and private score both 0.70</p>",
      "votes": null,
      "replies": [
        {
          "id": 726298,
          "author_name": "guozhiyu0914",
          "author_url": "",
          "post_date": "01/23/2020 00:50:13",
          "content": "<p>What base model have you chosen? I mean BERT or RoBERTa and...</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 726305,
          "author_name": "tonyxu",
          "author_url": "",
          "post_date": "01/23/2020 00:55:44",
          "content": "<p>I tried bert, xlnet, spanbert and found WWM BERT Large is the best.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 726293,
      "author_name": "guozhiyu0914",
      "author_url": "",
      "post_date": "01/23/2020 00:47:57",
      "content": "<p>I have chosen RoBERTa-large, I got 0.67, because of the kernel time limit, I haven't do ensembling.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 726329,
      "author_name": "naivelamb",
      "author_url": "",
      "post_date": "01/23/2020 01:12:34",
      "content": "<p>I have a bert-large with private lb 0.65 using TF2.0. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 726358,
      "author_name": "kentaronakanishi",
      "author_url": "",
      "post_date": "01/23/2020 01:25:46",
      "content": "<p>my best (TF2.0): Bert large (WWM) with last 2 layers combination (but I didn't use this because of public score)</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F193592%2F3ae56be4d3be988891a0a12fd2bb8cda%2F2020-01-23%2010.22.00.png?generation=1579742652969009&amp;alt=media\" alt=\"\"></p>\n\n<p>I'm surprised that some get over 0.67 in single model...!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 727032,
      "author_name": "seesee",
      "author_url": "",
      "post_date": "01/23/2020 11:55:40",
      "content": "<p>0.71 private lb: <a href=\"https://www.kaggle.com/seesee/submit-full\">https://www.kaggle.com/seesee/submit-full</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 727035,
      "author_name": "axel81",
      "author_url": "",
      "post_date": "01/23/2020 11:57:59",
      "content": "<p>0.64 private LB. Are you entering your best private LB score even if you didn't select it as your final submission?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "726264": "our best is 0.64 private lb https://www.kaggle.com/jiweiliu/jb-tf-2-0",
    "726279": "We have several single models at 0.69.  But they are pytorch ;)",
    "726284": "Have you implemented joint approach or?",
    "726286": "Variants of it yes.",
    "726291": "single tf1.0 model with public score and private score both 0.70",
    "726293": "I have chosen RoBERTa-large, I got 0.67, because of the kernel time limit, I haven't do ensembling.",
    "726298": "What base model have you chosen? I mean BERT or RoBERTa and...",
    "726305": "I tried bert, xlnet, spanbert and found WWM BERT Large is the best.",
    "726329": "I have a bert-large with private lb 0.65 using TF2.0.",
    "726358": "my best (TF2.0): Bert large (WWM) with last 2 layers combination (but I didn't use this because of public score)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F193592%2F3ae56be4d3be988891a0a12fd2bb8cda%2F2020-01-23%2010.22.00.png?generation=1579742652969009&amp;alt=media)\n\nI'm surprised that some get over 0.67 in single model...!",
    "726380": "cpmpml - Congrats :-) . Jus want to know how many single models you have used for ensembling. What was your approach for ensembling?",
    "727032": "0.71 private lb: https://www.kaggle.com/seesee/submit-full",
    "727035": "0.64 private LB. Are you entering your best private LB score even if you didn't select it as your final submission?",
    "727049": "s4sarath you can check in our summary :D",
    "727104": "Thank you :-)"
  },
  "source": "meta"
}