{
  "id": 107940,
  "title": "84th place solution",
  "url": "/competitions/aptos2019-blindness-detection/discussion/107940",
  "author_name": "Grigorev Artur",
  "post_date": "2019-09-08T03:40:27.042000",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Congrats to all the winners!</p>\n\n<p>Out solution is a linear combination of:</p>\n\n<ul>\n<li><p>2 x EfficientNetB3 trained purely on 2019 train data (with polar unrolling) for classification.\nImage size: 300x768. Score: 0.788, 0.782\n<a href=\"https://www.kaggle.com/agscin/polar-unrolling-preprocessing-84th-place\">https://www.kaggle.com/agscin/polar-unrolling-preprocessing-84th-place</a></p></li>\n<li><p>1 x EfficientNetB3 pretrained on 2015 data and fine-tuned on 2019 (with polar unrolling) for classification. Score: 0.794\nImage size: 300x768</p></li>\n</ul>\n\n<p>Training: Adam, 8 epochs 1e-3, then 8 epochs 1e-4.\nPreprocessing: autocrop, Ben Graham's preprocessing.\nImage augmentation: random zoom (1.0-1.1), random horizontal/vertical flip. Since the polar unrolling is used the is no need in rotation augmentation.</p>\n\n<p>TTA (test time augmentation): geometric mean of softmax predictions for horizontally/vertically flipped images.</p>\n\n<ul>\n<li>5 x ResNet50 trained for classification.\nImage size: 300x300</li>\n</ul>\n\n<p>Preprocessing: crop _ image _ from _ gray from Neuron Engineer's kernel.</p>",
  "messages": [
    {
      "id": 620890,
      "postDate": "2019-09-08T03:40:27.043Z",
      "content": "<p>Congrats to all the winners!</p>\n\n<p>Out solution is a linear combination of:</p>\n\n<ul>\n<li><p>2 x EfficientNetB3 trained purely on 2019 train data (with polar unrolling) for classification.\nImage size: 300x768. Score: 0.788, 0.782\n<a href=\"https://www.kaggle.com/agscin/polar-unrolling-preprocessing-84th-place\">https://www.kaggle.com/agscin/polar-unrolling-preprocessing-84th-place</a></p></li>\n<li><p>1 x EfficientNetB3 pretrained on 2015 data and fine-tuned on 2019 (with polar unrolling) for classification. Score: 0.794\nImage size: 300x768</p></li>\n</ul>\n\n<p>Training: Adam, 8 epochs 1e-3, then 8 epochs 1e-4.\nPreprocessing: autocrop, Ben Graham's preprocessing.\nImage augmentation: random zoom (1.0-1.1), random horizontal/vertical flip. Since the polar unrolling is used the is no need in rotation augmentation.</p>\n\n<p>TTA (test time augmentation): geometric mean of softmax predictions for horizontally/vertically flipped images.</p>\n\n<ul>\n<li>5 x ResNet50 trained for classification.\nImage size: 300x300</li>\n</ul>\n\n<p>Preprocessing: crop _ image _ from _ gray from Neuron Engineer's kernel.</p>",
      "rawMarkdown": "Congrats to all the winners!\n\nOut solution is a linear combination of:\n\n- 2 x EfficientNetB3 trained purely on 2019 train data (with polar unrolling) for classification.\nImage size: 300x768. Score: 0.788, 0.782\nhttps://www.kaggle.com/agscin/polar-unrolling-preprocessing-84th-place\n\n- 1 x EfficientNetB3 pretrained on 2015 data and fine-tuned on 2019 (with polar unrolling) for classification. Score: 0.794\nImage size: 300x768\n\nTraining: Adam, 8 epochs 1e-3, then 8 epochs 1e-4.\nPreprocessing: autocrop, Ben Graham's preprocessing.\nImage augmentation: random zoom (1.0-1.1), random horizontal/vertical flip. Since the polar unrolling is used the is no need in rotation augmentation.\n\nTTA (test time augmentation): geometric mean of softmax predictions for horizontally/vertically flipped images.\n\n- 5 x ResNet50 trained for classification.\nImage size: 300x300\n\nPreprocessing: crop _ image _ from _ gray from Neuron Engineer's kernel.\n\n\n",
      "votes": 4
    },
    {
      "id": 620900,
      "postDate": "2019-09-08T03:59:25.433Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 620900,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-09-08T03:59:25.433000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "620890": "Congrats to all the winners!\n\nOut solution is a linear combination of:\n\n- 2 x EfficientNetB3 trained purely on 2019 train data (with polar unrolling) for classification.\nImage size: 300x768. Score: 0.788, 0.782\nhttps://www.kaggle.com/agscin/polar-unrolling-preprocessing-84th-place\n\n- 1 x EfficientNetB3 pretrained on 2015 data and fine-tuned on 2019 (with polar unrolling) for classification. Score: 0.794\nImage size: 300x768\n\nTraining: Adam, 8 epochs 1e-3, then 8 epochs 1e-4.\nPreprocessing: autocrop, Ben Graham's preprocessing.\nImage augmentation: random zoom (1.0-1.1), random horizontal/vertical flip. Since the polar unrolling is used the is no need in rotation augmentation.\n\nTTA (test time augmentation): geometric mean of softmax predictions for horizontally/vertically flipped images.\n\n- 5 x ResNet50 trained for classification.\nImage size: 300x300\n\nPreprocessing: crop _ image _ from _ gray from Neuron Engineer's kernel.\n\n\n",
    "620900": ""
  }
}