{
  "id": 314524,
  "title": "1.00000 Private Leaderboard - Summary",
  "url": "/competitions/dog-vs-cat-classification/writeups/nmt-hi-u-1-00000-private-leaderboard-summary",
  "author_name": "",
  "post_date": "2022-03-29T12:30:04.153Z",
  "votes": 12,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Congrats to all the participants, and thanks to the host DataScience hosted such an interesting competetion.</p>\n<p>Let's start!</p>\n<p><strong>SETTING</strong><br>\nUsing Tensorflow<br>\nTraining using TPU on Google Colab Pro and Kaggle<br>\nBATCH SIZE = 256 and 512<br>\nclasses = ['dog', 'cat']<br>\noptimizer = AdamW(weight_decay=1e-8) with cosine_annealing_with_warmup<br>\nloss = CategoricalCrossentropy(label_smoothing=0.1)<br>\nmetrics=['accuracy', f1score]</p>\n<p><strong>AUGMENTATION</strong><br>\nRandomly rotate, shear, zoom, and shift<br>\nCutmix<br>\nMixup<br>\nCutout<br>\nGrid mask (<a href=\"https://www.kaggle.com/code/xiejialun/gridmask-data-augmentation-with-tensorflow/notebook\" target=\"_blank\">https://www.kaggle.com/code/xiejialun/gridmask-data-augmentation-with-tensorflow/notebook</a>)</p>\n<p><strong>STAGE 1</strong><br>\nSplitting training data by StratifiedKFold with n_fold = 5<br>\nTraining 5 EfficientnetV2S - img size: 224 and 5 EfficientnetV2L - img size: 288 models<br>\nUse these model to Pseudo Label on Test dataset:<br>\nImages with confidence &gt; 0.942 =&gt; A<br>\nImages with confidence &lt;= 0.942 and &gt; 0.93 =&gt; B<br>\nImages with confidence &lt;= 0.93 =&gt; C</p>\n<p><strong>STAGE 2</strong><br>\nTrain dataset = Original Train dataset + A<br>\nValid dataset = B<br>\nTrain 1 EfficientnetV2S - img size: 224</p>\n<p>EfficientnetV2S - img size: 224:<br>\nTotal params: 20,333,922<br>\nTrainable params: 20,180,050<br>\nNon-trainable params: 153,872</p>\n<p>Use this model to inference the Test set<br>\n=&gt; Final Solution</p>\n<p>Overall, i think i have used \"overkill\" solution because this is a good, balanced, simple dataset.<br>\nBut thanks to this competition, i could try many new and fun methods 😎.</p>\n<p><strong>LINK NOTEBOOK</strong><br>\n<a href=\"https://www.kaggle.com/code/v1olet5/dogvscat-1-0-private-solution-single-model-effv2s\" target=\"_blank\">https://www.kaggle.com/code/v1olet5/dogvscat-1-0-private-solution-single-model-effv2s</a><br>\n<a href=\"https://www.kaggle.com/code/v1olet5/3rd-nmt-hi-u-efficientnetv2s-20333922/notebook\" target=\"_blank\">https://www.kaggle.com/code/v1olet5/3rd-nmt-hi-u-efficientnetv2s-20333922/notebook</a></p>",
  "messages": [
    {
      "id": "1732116",
      "postDate": "03/23/2022 03:09:17",
      "content": "<p>Congrats to all the participants, and thanks to the host DataScience hosted such an interesting competetion.</p>\n<p>Let's start!</p>\n<p><strong>SETTING</strong><br>\nUsing Tensorflow<br>\nTraining using TPU on Google Colab Pro and Kaggle<br>\nBATCH SIZE = 256 and 512<br>\nclasses = ['dog', 'cat']<br>\noptimizer = AdamW(weight_decay=1e-8) with cosine_annealing_with_warmup<br>\nloss = CategoricalCrossentropy(label_smoothing=0.1)<br>\nmetrics=['accuracy', f1score]</p>\n<p><strong>AUGMENTATION</strong><br>\nRandomly rotate, shear, zoom, and shift<br>\nCutmix<br>\nMixup<br>\nCutout<br>\nGrid mask (<a href=\"https://www.kaggle.com/code/xiejialun/gridmask-data-augmentation-with-tensorflow/notebook\" target=\"_blank\">https://www.kaggle.com/code/xiejialun/gridmask-data-augmentation-with-tensorflow/notebook</a>)</p>\n<p><strong>STAGE 1</strong><br>\nSplitting training data by StratifiedKFold with n_fold = 5<br>\nTraining 5 EfficientnetV2S - img size: 224 and 5 EfficientnetV2L - img size: 288 models<br>\nUse these model to Pseudo Label on Test dataset:<br>\nImages with confidence &gt; 0.942 =&gt; A<br>\nImages with confidence &lt;= 0.942 and &gt; 0.93 =&gt; B<br>\nImages with confidence &lt;= 0.93 =&gt; C</p>\n<p><strong>STAGE 2</strong><br>\nTrain dataset = Original Train dataset + A<br>\nValid dataset = B<br>\nTrain 1 EfficientnetV2S - img size: 224</p>\n<p>EfficientnetV2S - img size: 224:<br>\nTotal params: 20,333,922<br>\nTrainable params: 20,180,050<br>\nNon-trainable params: 153,872</p>\n<p>Use this model to inference the Test set<br>\n=&gt; Final Solution</p>\n<p>Overall, i think i have used \"overkill\" solution because this is a good, balanced, simple dataset.<br>\nBut thanks to this competition, i could try many new and fun methods 😎.</p>\n<p><strong>LINK NOTEBOOK</strong><br>\n<a href=\"https://www.kaggle.com/code/v1olet5/dogvscat-1-0-private-solution-single-model-effv2s\" target=\"_blank\">https://www.kaggle.com/code/v1olet5/dogvscat-1-0-private-solution-single-model-effv2s</a><br>\n<a href=\"https://www.kaggle.com/code/v1olet5/3rd-nmt-hi-u-efficientnetv2s-20333922/notebook\" target=\"_blank\">https://www.kaggle.com/code/v1olet5/3rd-nmt-hi-u-efficientnetv2s-20333922/notebook</a></p>",
      "rawMarkdown": "Congrats to all the participants, and thanks to the host DataScience hosted such an interesting competetion.\n\nLet's start!\n\n**SETTING**\nUsing Tensorflow\nTraining using TPU on Google Colab Pro and Kaggle\nBATCH SIZE = 256 and 512\nclasses = ['dog', 'cat']\noptimizer = AdamW(weight_decay=1e-8) with cosine_annealing_with_warmup\nloss = CategoricalCrossentropy(label_smoothing=0.1)\nmetrics=['accuracy', f1score]\n\n**AUGMENTATION**\nRandomly rotate, shear, zoom, and shift\nCutmix\nMixup\nCutout\nGrid mask (https://www.kaggle.com/code/xiejialun/gridmask-data-augmentation-with-tensorflow/notebook)\n\n**STAGE 1**\nSplitting training data by StratifiedKFold with n_fold = 5\nTraining 5 EfficientnetV2S - img size: 224 and 5 EfficientnetV2L - img size: 288 models\nUse these model to Pseudo Label on Test dataset:\nImages with confidence > 0.942 => A\nImages with confidence <= 0.942 and > 0.93 => B\nImages with confidence <= 0.93 => C\n\n**STAGE 2**\nTrain dataset = Original Train dataset + A\nValid dataset = B\nTrain 1 EfficientnetV2S - img size: 224\n\nEfficientnetV2S - img size: 224:\nTotal params: 20,333,922\nTrainable params: 20,180,050\nNon-trainable params: 153,872\n\nUse this model to inference the Test set\n=> Final Solution\n\nOverall, i think i have used \"overkill\" solution because this is a good, balanced, simple dataset.\nBut thanks to this competition, i could try many new and fun methods 😎.\n\n**LINK NOTEBOOK**\nhttps://www.kaggle.com/code/v1olet5/dogvscat-1-0-private-solution-single-model-effv2s\nhttps://www.kaggle.com/code/v1olet5/3rd-nmt-hi-u-efficientnetv2s-20333922/notebook",
      "votes": null
    },
    {
      "id": "1735372",
      "postDate": "03/26/2022 06:54:38",
      "content": "<p>Great work! Thanks for such a creative approach!</p>",
      "rawMarkdown": "Great work! Thanks for such a creative approach!",
      "votes": null
    },
    {
      "id": "1735377",
      "postDate": "03/26/2022 07:02:31",
      "content": "<p>Thank you 👀</p>",
      "rawMarkdown": "Thank you 👀",
      "votes": null
    },
    {
      "id": "1814570",
      "postDate": "06/08/2022 04:13:29",
      "content": "<p>Seems like the links are all dead. Could you please update or send me via email <a>mrlasdt@gmail.com</a><br>\nBtw it's really nice solution, I tried EfficinetV2 too but that did not work :v</p>",
      "rawMarkdown": "Seems like the links are all dead. Could you please update or send me via email mrlasdt@gmail.com\nBtw it's really nice solution, I tried EfficinetV2 too but that did not work :v",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1735372,
      "author_name": "quangbinh",
      "author_url": "",
      "post_date": "03/26/2022 06:54:38",
      "content": "<p>Great work! Thanks for such a creative approach!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1735377,
          "author_name": "",
          "author_url": "",
          "post_date": "03/26/2022 07:02:31",
          "content": "<p>Thank you 👀</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1814570,
      "author_name": "willbi",
      "author_url": "",
      "post_date": "06/08/2022 04:13:29",
      "content": "<p>Seems like the links are all dead. Could you please update or send me via email <a>mrlasdt@gmail.com</a><br>\nBtw it's really nice solution, I tried EfficinetV2 too but that did not work :v</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1732116": "Congrats to all the participants, and thanks to the host DataScience hosted such an interesting competetion.\n\nLet's start!\n\n**SETTING**\nUsing Tensorflow\nTraining using TPU on Google Colab Pro and Kaggle\nBATCH SIZE = 256 and 512\nclasses = ['dog', 'cat']\noptimizer = AdamW(weight_decay=1e-8) with cosine_annealing_with_warmup\nloss = CategoricalCrossentropy(label_smoothing=0.1)\nmetrics=['accuracy', f1score]\n\n**AUGMENTATION**\nRandomly rotate, shear, zoom, and shift\nCutmix\nMixup\nCutout\nGrid mask (https://www.kaggle.com/code/xiejialun/gridmask-data-augmentation-with-tensorflow/notebook)\n\n**STAGE 1**\nSplitting training data by StratifiedKFold with n_fold = 5\nTraining 5 EfficientnetV2S - img size: 224 and 5 EfficientnetV2L - img size: 288 models\nUse these model to Pseudo Label on Test dataset:\nImages with confidence > 0.942 => A\nImages with confidence <= 0.942 and > 0.93 => B\nImages with confidence <= 0.93 => C\n\n**STAGE 2**\nTrain dataset = Original Train dataset + A\nValid dataset = B\nTrain 1 EfficientnetV2S - img size: 224\n\nEfficientnetV2S - img size: 224:\nTotal params: 20,333,922\nTrainable params: 20,180,050\nNon-trainable params: 153,872\n\nUse this model to inference the Test set\n=> Final Solution\n\nOverall, i think i have used \"overkill\" solution because this is a good, balanced, simple dataset.\nBut thanks to this competition, i could try many new and fun methods 😎.\n\n**LINK NOTEBOOK**\nhttps://www.kaggle.com/code/v1olet5/dogvscat-1-0-private-solution-single-model-effv2s\nhttps://www.kaggle.com/code/v1olet5/3rd-nmt-hi-u-efficientnetv2s-20333922/notebook",
    "1735372": "Great work! Thanks for such a creative approach!",
    "1735377": "Thank you 👀",
    "1814570": "Seems like the links are all dead. Could you please update or send me via email mrlasdt@gmail.com\nBtw it's really nice solution, I tried EfficinetV2 too but that did not work :v"
  },
  "source": "existing_discussion",
  "complete": true,
  "competition_id": 34426,
  "competition_slug": "dog-vs-cat-classification",
  "team_id": 8212655,
  "solution_url": "/competitions/dog-vs-cat-classification/writeups/nmt-hi-u-1-00000-private-leaderboard-summary"
}