{
  "id": 157513,
  "title": "My TPU solution - Private LB 0.96704",
  "url": "/competitions/flower-classification-with-tpus/writeups/bruce-young-my-tpu-solution-private-lb-0-96704",
  "author_name": "",
  "post_date": "2020-06-10T23:59:08.401892400Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>As the leaderboard is \"a bit\" messed up, I have no real idea where I ended up.\nHowever, this is what I did.</p>\n\n<p><a href=\"https://www.kaggle.com/mutantspore/flower-classification-petals-2a?scriptVersionId=33591653\">Flower Classification: Petals 2a</a>\nOnly the supplied competition data was used in this submission.</p>\n\n<p>Like most of us, I tried a bazillion different things but it turns out that my best Private Leaderboard score (from my 2 legal submissions) was a pretty simple affair. I started out, as most of us did with <strong>Martin Görner's</strong> - <a href=\"https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu\">Getting started with 100+ flowers on TPU</a>.</p>\n\n<p>I constructed an ensemble of EfficientNet B7, B5 and B4 using noisy-student weights which ran for 30 Epochs. I have another version, that I didn't submit, which ran for 35 epochs resulting in a slightly higher Private LB score (0.96753).</p>\n\n<p>The trick was to get as much processing in before the 3 hrs max TPU usage time was up.\nIt was a great learning experience that has helped me do well in the Plant Pathology and now the ongoing Melanoma competitions.</p>\n\n<p>thanks\nBruce</p>",
  "messages": [
    {
      "id": "881336",
      "postDate": "06/10/2020 23:59:08",
      "content": "<p>As the leaderboard is \"a bit\" messed up, I have no real idea where I ended up.\nHowever, this is what I did.</p>\n\n<p><a href=\"https://www.kaggle.com/mutantspore/flower-classification-petals-2a?scriptVersionId=33591653\">Flower Classification: Petals 2a</a>\nOnly the supplied competition data was used in this submission.</p>\n\n<p>Like most of us, I tried a bazillion different things but it turns out that my best Private Leaderboard score (from my 2 legal submissions) was a pretty simple affair. I started out, as most of us did with <strong>Martin Görner's</strong> - <a href=\"https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu\">Getting started with 100+ flowers on TPU</a>.</p>\n\n<p>I constructed an ensemble of EfficientNet B7, B5 and B4 using noisy-student weights which ran for 30 Epochs. I have another version, that I didn't submit, which ran for 35 epochs resulting in a slightly higher Private LB score (0.96753).</p>\n\n<p>The trick was to get as much processing in before the 3 hrs max TPU usage time was up.\nIt was a great learning experience that has helped me do well in the Plant Pathology and now the ongoing Melanoma competitions.</p>\n\n<p>thanks\nBruce</p>",
      "rawMarkdown": "As the leaderboard is \"a bit\" messed up, I have no real idea where I ended up.\nHowever, this is what I did.\n\n[Flower Classification: Petals 2a](https://www.kaggle.com/mutantspore/flower-classification-petals-2a?scriptVersionId=33591653)\nOnly the supplied competition data was used in this submission.\n\nLike most of us, I tried a bazillion different things but it turns out that my best Private Leaderboard score (from my 2 legal submissions) was a pretty simple affair. I started out, as most of us did with **Martin Görner's** - [Getting started with 100+ flowers on TPU](https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu).\n\nI constructed an ensemble of EfficientNet B7, B5 and B4 using noisy-student weights which ran for 30 Epochs. I have another version, that I didn't submit, which ran for 35 epochs resulting in a slightly higher Private LB score (0.96753).\n\nThe trick was to get as much processing in before the 3 hrs max TPU usage time was up.\nIt was a great learning experience that has helped me do well in the Plant Pathology and now the ongoing Melanoma competitions.\n\nthanks\nBruce",
      "votes": null
    },
    {
      "id": "882306",
      "postDate": "06/11/2020 17:53:36",
      "content": "<p>Great Job!</p>",
      "rawMarkdown": "Great Job!",
      "votes": null
    },
    {
      "id": "888468",
      "postDate": "06/16/2020 11:30:28",
      "content": "<p>Contrats Bruce, Great Job!</p>",
      "rawMarkdown": "Contrats Bruce, Great Job!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 882306,
      "author_name": "fuatsezer",
      "author_url": "",
      "post_date": "06/11/2020 17:53:36",
      "content": "<p>Great Job!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 888468,
      "author_name": "coreacasa",
      "author_url": "",
      "post_date": "06/16/2020 11:30:28",
      "content": "<p>Contrats Bruce, Great Job!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "881336": "As the leaderboard is \"a bit\" messed up, I have no real idea where I ended up.\nHowever, this is what I did.\n\n[Flower Classification: Petals 2a](https://www.kaggle.com/mutantspore/flower-classification-petals-2a?scriptVersionId=33591653)\nOnly the supplied competition data was used in this submission.\n\nLike most of us, I tried a bazillion different things but it turns out that my best Private Leaderboard score (from my 2 legal submissions) was a pretty simple affair. I started out, as most of us did with **Martin Görner's** - [Getting started with 100+ flowers on TPU](https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu).\n\nI constructed an ensemble of EfficientNet B7, B5 and B4 using noisy-student weights which ran for 30 Epochs. I have another version, that I didn't submit, which ran for 35 epochs resulting in a slightly higher Private LB score (0.96753).\n\nThe trick was to get as much processing in before the 3 hrs max TPU usage time was up.\nIt was a great learning experience that has helped me do well in the Plant Pathology and now the ongoing Melanoma competitions.\n\nthanks\nBruce",
    "882306": "Great Job!",
    "888468": "Contrats Bruce, Great Job!"
  },
  "source": "meta"
}