{
  "id": 235202,
  "title": "Images 384x384 in TFRecord format",
  "url": "/competitions/herbarium-2021-fgvc8/discussion/235202",
  "author_name": "Luigi Saetta",
  "post_date": "2021-04-28T09:21:13.847000",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>For those interested, I have published also the datasets with images 384x384, in TFRecord file format.<br>\nSince the size of the entire training set Is big, I have splitted in two datasets:</p>\n<p><a href=\"https://www.kaggle.com/luigisaetta/herb2021-384-1\" target=\"_blank\">Train1</a><br>\n<a href=\"https://www.kaggle.com/luigisaetta/herb2021-384-2\" target=\"_blank\">Train2</a><br>\n<a href=\"https://www.kaggle.com/luigisaetta/herb2021-test-384\" target=\"_blank\">Test</a></p>\n<p>Using a two-set split, you can easily start developing the train Notebook on a subset, and after scale to the entire dataset.</p>\n<p>Consider that the entire training, for enough epochs, on TPU takes 5 hours.</p>",
  "messages": [
    {
      "id": 1286667,
      "postDate": "2021-04-28T09:21:13.847Z",
      "content": "<p>For those interested, I have published also the datasets with images 384x384, in TFRecord file format.<br>\nSince the size of the entire training set Is big, I have splitted in two datasets:</p>\n<p><a href=\"https://www.kaggle.com/luigisaetta/herb2021-384-1\" target=\"_blank\">Train1</a><br>\n<a href=\"https://www.kaggle.com/luigisaetta/herb2021-384-2\" target=\"_blank\">Train2</a><br>\n<a href=\"https://www.kaggle.com/luigisaetta/herb2021-test-384\" target=\"_blank\">Test</a></p>\n<p>Using a two-set split, you can easily start developing the train Notebook on a subset, and after scale to the entire dataset.</p>\n<p>Consider that the entire training, for enough epochs, on TPU takes 5 hours.</p>",
      "rawMarkdown": "For those interested, I have published also the datasets with images 384x384, in TFRecord file format.\nSince the size of the entire training set Is big, I have splitted in two datasets:\n\n[Train1](https://www.kaggle.com/luigisaetta/herb2021-384-1)\n[Train2](https://www.kaggle.com/luigisaetta/herb2021-384-2)\n[Test](https://www.kaggle.com/luigisaetta/herb2021-test-384)\n\nUsing a two-set split, you can easily start developing the train Notebook on a subset, and after scale to the entire dataset.\n\nConsider that the entire training, for enough epochs, on TPU takes 5 hours.",
      "votes": 3
    },
    {
      "id": 1304748,
      "postDate": "2021-05-12T20:41:54.650Z",
      "content": "<p>I am experiencing  socket closed error while model.fit<br>\nI tried resizing data to smaller size and trying smaller model (efficientnetb2)</p>\n<p>did you also face this error ?<br>\nand how to solve it ?</p>\n<p>Thanks :)</p>",
      "rawMarkdown": "I am experiencing  socket closed error while model.fit\nI tried resizing data to smaller size and trying smaller model (efficientnetb2)\n\ndid you also face this error ?\nand how to solve it ?\n\nThanks :)\n\n\n",
      "votes": 1,
      "replies": [
        {
          "id": 1305210,
          "postDate": "2021-05-13T06:48:04.563Z",
          "content": "<p>You need to work with a smaller batch_size. on TPU I use 64…. I don't think it is possible to work on GPU…</p>",
          "rawMarkdown": "You need to work with a smaller batch_size. on TPU I use 64.... I don't think it is possible to work on GPU..."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1304748,
      "author_name": "Sagar",
      "author_url": "",
      "post_date": "2021-05-12T20:41:54.650000",
      "content": "<p>I am experiencing  socket closed error while model.fit<br>\nI tried resizing data to smaller size and trying smaller model (efficientnetb2)</p>\n<p>did you also face this error ?<br>\nand how to solve it ?</p>\n<p>Thanks :)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1305210,
          "author_name": "Luigi Saetta",
          "author_url": "",
          "post_date": "2021-05-13T06:48:04.563000",
          "content": "<p>You need to work with a smaller batch_size. on TPU I use 64…. I don't think it is possible to work on GPU…</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1286667": "For those interested, I have published also the datasets with images 384x384, in TFRecord file format.\nSince the size of the entire training set Is big, I have splitted in two datasets:\n\n[Train1](https://www.kaggle.com/luigisaetta/herb2021-384-1)\n[Train2](https://www.kaggle.com/luigisaetta/herb2021-384-2)\n[Test](https://www.kaggle.com/luigisaetta/herb2021-test-384)\n\nUsing a two-set split, you can easily start developing the train Notebook on a subset, and after scale to the entire dataset.\n\nConsider that the entire training, for enough epochs, on TPU takes 5 hours.",
    "1304748": "I am experiencing  socket closed error while model.fit\nI tried resizing data to smaller size and trying smaller model (efficientnetb2)\n\ndid you also face this error ?\nand how to solve it ?\n\nThanks :)\n\n\n"
  }
}