{
  "id": 40152,
  "title": "My solution with code Public 99.71  but Private 99.67  ",
  "url": "/competitions/carvana-image-masking-challenge/writeups/fujisan-my-solution-with-code-public-99-71-but-pri",
  "author_name": "",
  "post_date": "2017-09-28T11:47:15.833493500Z",
  "votes": 10,
  "comment_count": 3,
  "views": 0,
  "content": "<ol>\n<li><strong>Framework</strong>: Tensorflow(Keras)</li>\n<li><strong>Network</strong>: customized Unet referred to Wide Residual Network, Squeeze-and-Excitation Network <a href=\"https://github.com/fujisanx/kaggle/tree/master/Segmentation\">https://github.com/fujisanx/kaggle/tree/master/Segmentation</a></li>\n<li><strong>Training</strong>: 640x960 (640x959 1px padding), rotation, horizontal flip, scaling, HSV change, ganma correction, random erase</li>\n<li><strong>Inference</strong>: 640x960 ( crop 640x959 )  up size cv2.INTER_CUBIC</li>\n<li><strong>Validation</strong>: one hold out 20% data  </li>\n<li><strong>Test set predictions</strong>: 3 model predictions ( a little different paramater NN ) with TTA (flips).</li>\n<li><strong>Hardware</strong>: GTX1080ti x 1</li>\n</ol>\n\n<p>Japanese blog\n<a href=\"http://zero-ai.hatenablog.com/entry/2017/09/28/200328\">http://zero-ai.hatenablog.com/entry/2017/09/28/200328</a></p>",
  "messages": [
    {
      "id": "225158",
      "postDate": "09/28/2017 11:47:15",
      "content": "<ol>\n<li><strong>Framework</strong>: Tensorflow(Keras)</li>\n<li><strong>Network</strong>: customized Unet referred to Wide Residual Network, Squeeze-and-Excitation Network <a href=\"https://github.com/fujisanx/kaggle/tree/master/Segmentation\">https://github.com/fujisanx/kaggle/tree/master/Segmentation</a></li>\n<li><strong>Training</strong>: 640x960 (640x959 1px padding), rotation, horizontal flip, scaling, HSV change, ganma correction, random erase</li>\n<li><strong>Inference</strong>: 640x960 ( crop 640x959 )  up size cv2.INTER_CUBIC</li>\n<li><strong>Validation</strong>: one hold out 20% data  </li>\n<li><strong>Test set predictions</strong>: 3 model predictions ( a little different paramater NN ) with TTA (flips).</li>\n<li><strong>Hardware</strong>: GTX1080ti x 1</li>\n</ol>\n\n<p>Japanese blog\n<a href=\"http://zero-ai.hatenablog.com/entry/2017/09/28/200328\">http://zero-ai.hatenablog.com/entry/2017/09/28/200328</a></p>",
      "rawMarkdown": "1. **Framework**: Tensorflow(Keras)\n 2. **Network**: customized Unet referred to Wide Residual Network, Squeeze-and-Excitation Network https://github.com/fujisanx/kaggle/tree/master/Segmentation\n 3. **Training**: 640x960 (640x959 1px padding), rotation, horizontal flip, scaling, HSV change, ganma correction, random erase\n 4. **Inference**: 640x960 ( crop 640x959 )  up size cv2.INTER_CUBIC\n 5. **Validation**: one hold out 20% data  \n 6. **Test set predictions**: 3 model predictions ( a little different paramater NN ) with TTA (flips).\n 7. **Hardware**: GTX1080ti x 1\n\nJapanese blog\nhttp://zero-ai.hatenablog.com/entry/2017/09/28/200328",
      "votes": null
    },
    {
      "id": "225162",
      "postDate": "09/28/2017 11:53:18",
      "content": "<p>So cool you use squeeze and excitation! I only learned about it yesterday!!!</p>",
      "rawMarkdown": "So cool you use squeeze and excitation! I only learned about it yesterday!!!",
      "votes": null
    },
    {
      "id": "225196",
      "postDate": "09/28/2017 13:29:30",
      "content": "<p>Awesome! You've got all the cool networks we didn't have time to try. </p>",
      "rawMarkdown": "Awesome! You've got all the cool networks we didn't have time to try.",
      "votes": null
    },
    {
      "id": "225364",
      "postDate": "09/28/2017 20:48:23",
      "content": "<p>@Fijusan thanks for sharing and congratulations.</p>",
      "rawMarkdown": "Fijusan thanks for sharing and congratulations.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 225162,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "09/28/2017 11:53:18",
      "content": "<p>So cool you use squeeze and excitation! I only learned about it yesterday!!!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225196,
      "author_name": "ceperaang",
      "author_url": "",
      "post_date": "09/28/2017 13:29:30",
      "content": "<p>Awesome! You've got all the cool networks we didn't have time to try. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 225364,
      "author_name": "sheriytm",
      "author_url": "",
      "post_date": "09/28/2017 20:48:23",
      "content": "<p>@Fijusan thanks for sharing and congratulations.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "225158": "1. **Framework**: Tensorflow(Keras)\n 2. **Network**: customized Unet referred to Wide Residual Network, Squeeze-and-Excitation Network https://github.com/fujisanx/kaggle/tree/master/Segmentation\n 3. **Training**: 640x960 (640x959 1px padding), rotation, horizontal flip, scaling, HSV change, ganma correction, random erase\n 4. **Inference**: 640x960 ( crop 640x959 )  up size cv2.INTER_CUBIC\n 5. **Validation**: one hold out 20% data  \n 6. **Test set predictions**: 3 model predictions ( a little different paramater NN ) with TTA (flips).\n 7. **Hardware**: GTX1080ti x 1\n\nJapanese blog\nhttp://zero-ai.hatenablog.com/entry/2017/09/28/200328",
    "225162": "So cool you use squeeze and excitation! I only learned about it yesterday!!!",
    "225196": "Awesome! You've got all the cool networks we didn't have time to try.",
    "225364": "Fijusan thanks for sharing and congratulations."
  },
  "source": "meta"
}