{
  "id": 234553,
  "title": "4h training on TPU for LB 5.3: good or inefficient? (LTSM+ATT)",
  "url": "/competitions/bms-molecular-translation/discussion/234553",
  "author_name": "",
  "post_date": "2021-04-24T22:41:28.224769900Z",
  "votes": 7,
  "comment_count": 9,
  "views": 0,
  "content": "<p>Hey guys,<br>\nI'm very curious for how long other TPU users train their model to reach a certain LB score.<br>\nI have an LTSM + attention model that reached LB 5.3 after 4h on the kaggle TPU. <br>\nIs that good / kinda normal or inefficient?<br>\nIs it worth putting in more time?</p>",
  "messages": [
    {
      "id": "1283418",
      "postDate": "04/24/2021 22:41:28",
      "content": "<p>Hey guys,<br>\nI'm very curious for how long other TPU users train their model to reach a certain LB score.<br>\nI have an LTSM + attention model that reached LB 5.3 after 4h on the kaggle TPU. <br>\nIs that good / kinda normal or inefficient?<br>\nIs it worth putting in more time?</p>",
      "rawMarkdown": "Hey guys,\nI'm very curious for how long other TPU users train their model to reach a certain LB score.\nI have an LTSM + attention model that reached LB 5.3 after 4h on the kaggle TPU. \nIs that good / kinda normal or inefficient?\nIs it worth putting in more time?",
      "votes": null
    },
    {
      "id": "1283459",
      "postDate": "04/25/2021 00:24:40",
      "content": "<p>\"Is it worth putting in more time?\"</p>\n<p>extrapolate the training loss. this can help you estimate the best results that you can get</p>",
      "rawMarkdown": "\"Is it worth putting in more time?\"\n\nextrapolate the training loss. this can help you estimate the best results that you can get",
      "votes": null
    },
    {
      "id": "1283482",
      "postDate": "04/25/2021 01:06:34",
      "content": "<p>From my experience after so many experiments on TPU with the LSTM+ATT model, I can clearly say your results are good and worth trying!</p>",
      "rawMarkdown": "From my experience after so many experiments on TPU with the LSTM+ATT model, I can clearly say your results are good and worth trying!",
      "votes": null
    },
    {
      "id": "1284082",
      "postDate": "04/25/2021 14:41:57",
      "content": "<p>Predicting the quality of future prediction results, got it :)<br>\nThanks</p>",
      "rawMarkdown": "Predicting the quality of future prediction results, got it :)\nThanks",
      "votes": null
    },
    {
      "id": "1284084",
      "postDate": "04/25/2021 14:42:56",
      "content": "<p>Great to hear, thanks</p>",
      "rawMarkdown": "Great to hear, thanks",
      "votes": null
    },
    {
      "id": "1287166",
      "postDate": "04/28/2021 18:46:26",
      "content": "<p>Which model in encoder you used to reach for LB 5.3? </p>",
      "rawMarkdown": "Which model in encoder you used to reach for LB 5.3?",
      "votes": null
    },
    {
      "id": "1287235",
      "postDate": "04/28/2021 20:20:30",
      "content": "<p>EfficientNet B2 with 512x256 resolution</p>",
      "rawMarkdown": "EfficientNet B2 with 512x256 resolution",
      "votes": null
    },
    {
      "id": "1287239",
      "postDate": "04/28/2021 20:24:32",
      "content": "<p>That's super cool. Now seem like you enter 4. ;) </p>",
      "rawMarkdown": "That's super cool. Now seem like you enter 4. ;)",
      "votes": null
    },
    {
      "id": "1287401",
      "postDate": "04/29/2021 03:50:10",
      "content": "<p>The resolution you mentioned is it Decoder Dim x Attention Units?</p>",
      "rawMarkdown": "The resolution you mentioned is it Decoder Dim x Attention Units?",
      "votes": null
    },
    {
      "id": "1287663",
      "postDate": "04/29/2021 09:38:49",
      "content": "<p>Actually I meant the image resolution, but the decoder dim x attention units is also 512x256</p>",
      "rawMarkdown": "Actually I meant the image resolution, but the decoder dim x attention units is also 512x256",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1283459,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "04/25/2021 00:24:40",
      "content": "<p>\"Is it worth putting in more time?\"</p>\n<p>extrapolate the training loss. this can help you estimate the best results that you can get</p>",
      "votes": null,
      "replies": [
        {
          "id": 1284082,
          "author_name": "michaelwolff",
          "author_url": "",
          "post_date": "04/25/2021 14:41:57",
          "content": "<p>Predicting the quality of future prediction results, got it :)<br>\nThanks</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1283482,
      "author_name": "sorkun",
      "author_url": "",
      "post_date": "04/25/2021 01:06:34",
      "content": "<p>From my experience after so many experiments on TPU with the LSTM+ATT model, I can clearly say your results are good and worth trying!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1284084,
          "author_name": "michaelwolff",
          "author_url": "",
          "post_date": "04/25/2021 14:42:56",
          "content": "<p>Great to hear, thanks</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1287166,
      "author_name": "aifahim",
      "author_url": "",
      "post_date": "04/28/2021 18:46:26",
      "content": "<p>Which model in encoder you used to reach for LB 5.3? </p>",
      "votes": null,
      "replies": [
        {
          "id": 1287235,
          "author_name": "michaelwolff",
          "author_url": "",
          "post_date": "04/28/2021 20:20:30",
          "content": "<p>EfficientNet B2 with 512x256 resolution</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1287239,
          "author_name": "aifahim",
          "author_url": "",
          "post_date": "04/28/2021 20:24:32",
          "content": "<p>That's super cool. Now seem like you enter 4. ;) </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1287401,
          "author_name": "aifahim",
          "author_url": "",
          "post_date": "04/29/2021 03:50:10",
          "content": "<p>The resolution you mentioned is it Decoder Dim x Attention Units?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1287663,
          "author_name": "michaelwolff",
          "author_url": "",
          "post_date": "04/29/2021 09:38:49",
          "content": "<p>Actually I meant the image resolution, but the decoder dim x attention units is also 512x256</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1283418": "Hey guys,\nI'm very curious for how long other TPU users train their model to reach a certain LB score.\nI have an LTSM + attention model that reached LB 5.3 after 4h on the kaggle TPU. \nIs that good / kinda normal or inefficient?\nIs it worth putting in more time?",
    "1283459": "\"Is it worth putting in more time?\"\n\nextrapolate the training loss. this can help you estimate the best results that you can get",
    "1283482": "From my experience after so many experiments on TPU with the LSTM+ATT model, I can clearly say your results are good and worth trying!",
    "1284082": "Predicting the quality of future prediction results, got it :)\nThanks",
    "1284084": "Great to hear, thanks",
    "1287166": "Which model in encoder you used to reach for LB 5.3?",
    "1287235": "EfficientNet B2 with 512x256 resolution",
    "1287239": "That's super cool. Now seem like you enter 4. ;)",
    "1287401": "The resolution you mentioned is it Decoder Dim x Attention Units?",
    "1287663": "Actually I meant the image resolution, but the decoder dim x attention units is also 512x256"
  },
  "source": "meta"
}