{
  "id": 358203,
  "title": "13th Place Solution",
  "url": "/competitions/mayo-clinic-strip-ai/discussion/358203",
  "author_name": "Darien Schettler",
  "post_date": "2022-10-07T01:08:42.491000",
  "votes": 24,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi there, this is a pretty big shock. I put together a solid initial baseline early on in this competition and got busy with other things. I open-sourced this approach <a href=\"https://www.kaggle.com/code/dschettler8845/13th-tf-mcsai-no-tiling-model-inference/notebook?scriptVersionId=100779671\" target=\"_blank\"><strong>inference here</strong></a> and <a href=\"https://www.kaggle.com/code/dschettler8845/mcsai-no-tiling-model-tpu-tf\" target=\"_blank\"><strong>training here</strong></a>.</p>\n<p>This notebook was created almost 3 months ago (within the first few weeks … maybe the first week … of the competition). </p>\n<p><br></p>\n<p><strong>The basics are as follows:</strong></p>\n<ul>\n<li>Pretrained EfficientNetB6 fine-tuned w/ TPU on WSI</li>\n<li>Model Head --&gt; Dropout @ 0.5 &gt; 2 Node CC w/ Class Weights</li>\n<li>Learning Rate Ramps Up and Decays (12 Epochs)</li>\n<li>Whole Slides Used (no tiling)</li>\n<li>Images are resized to (512,512,3)</li>\n<li>Use pyvips for WSI processing</li>\n<li>Train Augmentation</li>\n</ul>\n<pre><code>def augment_batch(img_batch):\n    img_batch = tf.image.random_brightness(img_batch, 0.2)\n    img_batch = tf.image.random_contrast(img_batch, 0.5, 2.0)\n    img_batch = tf.image.random_saturation(img_batch, 0.75, 1.25)\n    img_batch = tf.image.random_hue(img_batch, 0.1)\n    return img_batch\n</code></pre>\n<p>If anyone has any questions please let me know! <br>\nAlso, thanks to Kaggle for constantly hosting interesting Biology competitions.</p>\n<p><br></p>\n<p>ps. This levelled me up to a Competitions Master. Woo!</p>",
  "messages": [
    {
      "id": 1975703,
      "postDate": "2022-10-07T01:08:42.493Z",
      "content": "<p>Hi there, this is a pretty big shock. I put together a solid initial baseline early on in this competition and got busy with other things. I open-sourced this approach <a href=\"https://www.kaggle.com/code/dschettler8845/13th-tf-mcsai-no-tiling-model-inference/notebook?scriptVersionId=100779671\" target=\"_blank\"><strong>inference here</strong></a> and <a href=\"https://www.kaggle.com/code/dschettler8845/mcsai-no-tiling-model-tpu-tf\" target=\"_blank\"><strong>training here</strong></a>.</p>\n<p>This notebook was created almost 3 months ago (within the first few weeks … maybe the first week … of the competition). </p>\n<p><br></p>\n<p><strong>The basics are as follows:</strong></p>\n<ul>\n<li>Pretrained EfficientNetB6 fine-tuned w/ TPU on WSI</li>\n<li>Model Head --&gt; Dropout @ 0.5 &gt; 2 Node CC w/ Class Weights</li>\n<li>Learning Rate Ramps Up and Decays (12 Epochs)</li>\n<li>Whole Slides Used (no tiling)</li>\n<li>Images are resized to (512,512,3)</li>\n<li>Use pyvips for WSI processing</li>\n<li>Train Augmentation</li>\n</ul>\n<pre><code>def augment_batch(img_batch):\n    img_batch = tf.image.random_brightness(img_batch, 0.2)\n    img_batch = tf.image.random_contrast(img_batch, 0.5, 2.0)\n    img_batch = tf.image.random_saturation(img_batch, 0.75, 1.25)\n    img_batch = tf.image.random_hue(img_batch, 0.1)\n    return img_batch\n</code></pre>\n<p>If anyone has any questions please let me know! <br>\nAlso, thanks to Kaggle for constantly hosting interesting Biology competitions.</p>\n<p><br></p>\n<p>ps. This levelled me up to a Competitions Master. Woo!</p>",
      "rawMarkdown": "Hi there, this is a pretty big shock. I put together a solid initial baseline early on in this competition and got busy with other things. I open-sourced this approach [**inference here**](https://www.kaggle.com/code/dschettler8845/13th-tf-mcsai-no-tiling-model-inference/notebook?scriptVersionId=100779671) and [**training here**](https://www.kaggle.com/code/dschettler8845/mcsai-no-tiling-model-tpu-tf).\n\nThis notebook was created almost 3 months ago (within the first few weeks ... maybe the first week ... of the competition). \n\n<br>\n\n**The basics are as follows:**\n* Pretrained EfficientNetB6 fine-tuned w/ TPU on WSI\n* Model Head --> Dropout @ 0.5 > 2 Node CC w/ Class Weights\n* Learning Rate Ramps Up and Decays (12 Epochs)\n* Whole Slides Used (no tiling)\n* Images are resized to (512,512,3)\n* Use pyvips for WSI processing\n* Train Augmentation\n```python\ndef augment_batch(img_batch):\n    img_batch = tf.image.random_brightness(img_batch, 0.2)\n    img_batch = tf.image.random_contrast(img_batch, 0.5, 2.0)\n    img_batch = tf.image.random_saturation(img_batch, 0.75, 1.25)\n    img_batch = tf.image.random_hue(img_batch, 0.1)\n    return img_batch\n```\n\nIf anyone has any questions please let me know! \nAlso, thanks to Kaggle for constantly hosting interesting Biology competitions.\n\n<br>\n\nps. This levelled me up to a Competitions Master. Woo!",
      "votes": 24
    },
    {
      "id": 1975901,
      "postDate": "2022-10-07T04:40:56.433Z",
      "content": "<p>Congratulations, and thank you for sharing your work!</p>",
      "rawMarkdown": "Congratulations, and thank you for sharing your work!",
      "votes": 1
    },
    {
      "id": 1975816,
      "postDate": "2022-10-07T03:24:15.567Z",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/dschettler8845\" target=\"_blank\">@dschettler8845</a>, and thanks for your early notebooks and suggestions. Those helped set the tone to tackle these huge images :)</p>",
      "rawMarkdown": "Congrats @dschettler8845, and thanks for your early notebooks and suggestions. Those helped set the tone to tackle these huge images :)",
      "votes": 1,
      "replies": [
        {
          "id": 1980192,
          "postDate": "2022-10-10T03:02:53.593Z",
          "content": "<p>You’re welcome! Congrats on 7th!</p>",
          "rawMarkdown": "You’re welcome! Congrats on 7th!",
          "votes": 1
        }
      ]
    },
    {
      "id": 2102552,
      "postDate": "2023-01-16T17:23:15.253Z",
      "content": "<p>Congratulations <a href=\"https://www.kaggle.com/dschettler8845\" target=\"_blank\">@dschettler8845</a>. For modeling and training which insights would you share on either Keras or Tensorflow or Yolo Fast-RNN, or other techniques would be instructive or appropriate to begin first level investigation for a similar problem in image classification. AKA, what learned lessons would you advise on modeling for performance on image classifications?</p>",
      "rawMarkdown": "Congratulations @dschettler8845. For modeling and training which insights would you share on either Keras or Tensorflow or Yolo Fast-RNN, or other techniques would be instructive or appropriate to begin first level investigation for a similar problem in image classification. AKA, what learned lessons would you advise on modeling for performance on image classifications?"
    },
    {
      "id": 2046585,
      "postDate": "2022-11-28T08:54:05.823Z",
      "content": "<p>Hey, congratulations on your work. I am super curious about how this model makes predictions ^^. Have you tried performing model explainability on this model? </p>",
      "rawMarkdown": "Hey, congratulations on your work. I am super curious about how this model makes predictions ^^. Have you tried performing model explainability on this model? "
    },
    {
      "id": 1983781,
      "postDate": "2022-10-12T08:24:17.770Z",
      "content": "<p>Thanks for uploading your work….lot's to learn from it.</p>",
      "rawMarkdown": "Thanks for uploading your work....lot's to learn from it."
    }
  ],
  "comments": [
    {
      "id": 1975901,
      "author_name": "Hassan Abedi",
      "author_url": "",
      "post_date": "2022-10-07T04:40:56.433000",
      "content": "<p>Congratulations, and thank you for sharing your work!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1975816,
      "author_name": "tdiceman",
      "author_url": "",
      "post_date": "2022-10-07T03:24:15.567000",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/dschettler8845\" target=\"_blank\">@dschettler8845</a>, and thanks for your early notebooks and suggestions. Those helped set the tone to tackle these huge images :)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1980192,
          "author_name": "Darien Schettler",
          "author_url": "",
          "post_date": "2022-10-10T03:02:53.593000",
          "content": "<p>You’re welcome! Congrats on 7th!</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 2102552,
      "author_name": "@500 matthew yeseta",
      "author_url": "",
      "post_date": "2023-01-16T17:23:15.253000",
      "content": "<p>Congratulations <a href=\"https://www.kaggle.com/dschettler8845\" target=\"_blank\">@dschettler8845</a>. For modeling and training which insights would you share on either Keras or Tensorflow or Yolo Fast-RNN, or other techniques would be instructive or appropriate to begin first level investigation for a similar problem in image classification. AKA, what learned lessons would you advise on modeling for performance on image classifications?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2046585,
      "author_name": "trunghjieu",
      "author_url": "",
      "post_date": "2022-11-28T08:54:05.823000",
      "content": "<p>Hey, congratulations on your work. I am super curious about how this model makes predictions ^^. Have you tried performing model explainability on this model? </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1983781,
      "author_name": "YugalKishore",
      "author_url": "",
      "post_date": "2022-10-12T08:24:17.770000",
      "content": "<p>Thanks for uploading your work….lot's to learn from it.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1975703": "Hi there, this is a pretty big shock. I put together a solid initial baseline early on in this competition and got busy with other things. I open-sourced this approach [**inference here**](https://www.kaggle.com/code/dschettler8845/13th-tf-mcsai-no-tiling-model-inference/notebook?scriptVersionId=100779671) and [**training here**](https://www.kaggle.com/code/dschettler8845/mcsai-no-tiling-model-tpu-tf).\n\nThis notebook was created almost 3 months ago (within the first few weeks ... maybe the first week ... of the competition). \n\n<br>\n\n**The basics are as follows:**\n* Pretrained EfficientNetB6 fine-tuned w/ TPU on WSI\n* Model Head --> Dropout @ 0.5 > 2 Node CC w/ Class Weights\n* Learning Rate Ramps Up and Decays (12 Epochs)\n* Whole Slides Used (no tiling)\n* Images are resized to (512,512,3)\n* Use pyvips for WSI processing\n* Train Augmentation\n```python\ndef augment_batch(img_batch):\n    img_batch = tf.image.random_brightness(img_batch, 0.2)\n    img_batch = tf.image.random_contrast(img_batch, 0.5, 2.0)\n    img_batch = tf.image.random_saturation(img_batch, 0.75, 1.25)\n    img_batch = tf.image.random_hue(img_batch, 0.1)\n    return img_batch\n```\n\nIf anyone has any questions please let me know! \nAlso, thanks to Kaggle for constantly hosting interesting Biology competitions.\n\n<br>\n\nps. This levelled me up to a Competitions Master. Woo!",
    "1975901": "Congratulations, and thank you for sharing your work!",
    "1975816": "Congrats @dschettler8845, and thanks for your early notebooks and suggestions. Those helped set the tone to tackle these huge images :)",
    "2102552": "Congratulations @dschettler8845. For modeling and training which insights would you share on either Keras or Tensorflow or Yolo Fast-RNN, or other techniques would be instructive or appropriate to begin first level investigation for a similar problem in image classification. AKA, what learned lessons would you advise on modeling for performance on image classifications?",
    "2046585": "Hey, congratulations on your work. I am super curious about how this model makes predictions ^^. Have you tried performing model explainability on this model? ",
    "1983781": "Thanks for uploading your work....lot's to learn from it."
  }
}