{
  "id": 159005,
  "title": "TPU prize solution (Private Lb : 0.96816)",
  "url": "/competitions/flower-classification-with-tpus/writeups/zenerdiode818-tpu-prize-solution-private-lb-0-9681",
  "author_name": "",
  "post_date": "2020-06-16T04:59:48.910662900Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I  thank Kaggle for organizing this competition. It was a great learning experience to work with TPU technology.</p>\n\n<p><strong>My Solution:</strong>\nI made an ensemble from EfficientNet B7(Noisy student weights) and Densenet 201. The former is trained for 32 epochs and the latter, 20 epochs. No data augmentations were used. Predictions were done using TTA technique.(<a href=\"https://www.kaggle.com/gskdhiman/enet-b7-densenet-with-tta/data\">https://www.kaggle.com/gskdhiman/enet-b7-densenet-with-tta/data</a>). Like many other kernels, training without validation provided higher score on public leaderboard.</p>\n\n<p>I wound like to give credits to the following notebooks:\n<a href=\"https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu\">https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu</a>\n<a href=\"https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96\">https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96</a></p>\n\n<p>My Notebook link:\n<a href=\"https://www.kaggle.com/zenerdiode818/flower-8-3-20/notebook\">https://www.kaggle.com/zenerdiode818/flower-8-3-20/notebook</a>\nP.S: Notebook is a bit clumsy and i can't edit it now.</p>\n\n<p>Thanks,\nZener Diode</p>",
  "messages": [
    {
      "id": "888053",
      "postDate": "06/16/2020 04:59:48",
      "content": "<p>I  thank Kaggle for organizing this competition. It was a great learning experience to work with TPU technology.</p>\n\n<p><strong>My Solution:</strong>\nI made an ensemble from EfficientNet B7(Noisy student weights) and Densenet 201. The former is trained for 32 epochs and the latter, 20 epochs. No data augmentations were used. Predictions were done using TTA technique.(<a href=\"https://www.kaggle.com/gskdhiman/enet-b7-densenet-with-tta/data\">https://www.kaggle.com/gskdhiman/enet-b7-densenet-with-tta/data</a>). Like many other kernels, training without validation provided higher score on public leaderboard.</p>\n\n<p>I wound like to give credits to the following notebooks:\n<a href=\"https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu\">https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu</a>\n<a href=\"https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96\">https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96</a></p>\n\n<p>My Notebook link:\n<a href=\"https://www.kaggle.com/zenerdiode818/flower-8-3-20/notebook\">https://www.kaggle.com/zenerdiode818/flower-8-3-20/notebook</a>\nP.S: Notebook is a bit clumsy and i can't edit it now.</p>\n\n<p>Thanks,\nZener Diode</p>",
      "rawMarkdown": "I  thank Kaggle for organizing this competition. It was a great learning experience to work with TPU technology.\n\n**My Solution:**\nI made an ensemble from EfficientNet B7(Noisy student weights) and Densenet 201. The former is trained for 32 epochs and the latter, 20 epochs. No data augmentations were used. Predictions were done using TTA technique.(https://www.kaggle.com/gskdhiman/enet-b7-densenet-with-tta/data). Like many other kernels, training without validation provided higher score on public leaderboard.\n\nI wound like to give credits to the following notebooks:\nhttps://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu\nhttps://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96\n\n\nMy Notebook link:\nhttps://www.kaggle.com/zenerdiode818/flower-8-3-20/notebook\nP.S: Notebook is a bit clumsy and i can't edit it now.\n\n\nThanks,\nZener Diode",
      "votes": null
    },
    {
      "id": "888469",
      "postDate": "06/16/2020 11:31:41",
      "content": "<p>Congrats for your Prize!</p>",
      "rawMarkdown": "Congrats for your Prize!",
      "votes": null
    },
    {
      "id": "888524",
      "postDate": "06/16/2020 12:21:30",
      "content": "<p>Thanks!! Same to you <a href=\"/coreacasa\">@coreacasa</a>  </p>",
      "rawMarkdown": "Thanks!! Same to you @coreacasa",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 888469,
      "author_name": "coreacasa",
      "author_url": "",
      "post_date": "06/16/2020 11:31:41",
      "content": "<p>Congrats for your Prize!</p>",
      "votes": null,
      "replies": [
        {
          "id": 888524,
          "author_name": "zenerdiode818",
          "author_url": "",
          "post_date": "06/16/2020 12:21:30",
          "content": "<p>Thanks!! Same to you <a href=\"/coreacasa\">@coreacasa</a>  </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "888053": "I  thank Kaggle for organizing this competition. It was a great learning experience to work with TPU technology.\n\n**My Solution:**\nI made an ensemble from EfficientNet B7(Noisy student weights) and Densenet 201. The former is trained for 32 epochs and the latter, 20 epochs. No data augmentations were used. Predictions were done using TTA technique.(https://www.kaggle.com/gskdhiman/enet-b7-densenet-with-tta/data). Like many other kernels, training without validation provided higher score on public leaderboard.\n\nI wound like to give credits to the following notebooks:\nhttps://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu\nhttps://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96\n\n\nMy Notebook link:\nhttps://www.kaggle.com/zenerdiode818/flower-8-3-20/notebook\nP.S: Notebook is a bit clumsy and i can't edit it now.\n\n\nThanks,\nZener Diode",
    "888469": "Congrats for your Prize!",
    "888524": "Thanks!! Same to you @coreacasa"
  },
  "source": "meta"
}