{
  "id": 106452,
  "title": "8th place solution(private 89)",
  "url": "/competitions/generative-dog-images/writeups/tkato-8th-place-solution-private-89",
  "author_name": "",
  "post_date": "2019-08-29T14:56:51.122351400Z",
  "votes": 14,
  "comment_count": 1,
  "views": 0,
  "content": "<h1>code</h1>\n\n<ul>\n<li><p>kernel\n<a href=\"https://www.kaggle.com/katotetsuro/stylegan-private-89?scriptVersionId=18524688\">https://www.kaggle.com/katotetsuro/stylegan-private-89?scriptVersionId=18524688</a></p></li>\n<li><p>github\n<a href=\"https://github.com/katotetsuro/chainer-stylegan/tree/ac\">https://github.com/katotetsuro/chainer-stylegan/tree/ac</a></p></li>\n</ul>\n\n<p>FYI: This scripting style is borrowed from this:\n<a href=\"https://github.com/lopuhin/kaggle-script-template\">https://github.com/lopuhin/kaggle-script-template</a></p>\n\n<p>I prefer script style, which is easy to manage by git.</p>\n\n<h1>summary</h1>\n\n<p>To be honest, there is no special technique in my solution.\nI chose stylegan, then looked for good parameters.\nMy stylegan starts with public score 140, then tuning parameters, I got 35.</p>\n\n<ul>\n<li>for preprocessing: crop image by bounding box, random horizontal flip</li>\n<li>reduce input channel size of style generator to 256 because of resource limitation.</li>\n</ul>\n\n<p>all hyper parameters are shown in below.\n<a href=\"https://github.com/katotetsuro/chainer-stylegan/blob/ac/src/stylegan/config.py\">https://github.com/katotetsuro/chainer-stylegan/blob/ac/src/stylegan/config.py</a>\n<a href=\"https://github.com/katotetsuro/chainer-stylegan/blob/ac/script_template.py\">https://github.com/katotetsuro/chainer-stylegan/blob/ac/script_template.py</a></p>\n\n<h1>trials that didn't work for me</h1>\n\n<ul>\n<li>Truncation trick from stylegan paper. I tried various truncation strength, but without truncation was the best.</li>\n<li>Auxiliary Classifier</li>\n<li>Data cleaning. I thought training data quality is quite important, so I did some procedure (dog face detection, manually picking good images..) but nothing worked.\nToday, I read some kernels and learned  there's some good techniques.\nThis is one of the room for improvement of my solution, I guess.</li>\n</ul>\n\n<p>It was really fun competition! Thanks for kaggle team!!</p>",
  "messages": [
    {
      "id": "612005",
      "postDate": "08/29/2019 14:56:51",
      "content": "<h1>code</h1>\n\n<ul>\n<li><p>kernel\n<a href=\"https://www.kaggle.com/katotetsuro/stylegan-private-89?scriptVersionId=18524688\">https://www.kaggle.com/katotetsuro/stylegan-private-89?scriptVersionId=18524688</a></p></li>\n<li><p>github\n<a href=\"https://github.com/katotetsuro/chainer-stylegan/tree/ac\">https://github.com/katotetsuro/chainer-stylegan/tree/ac</a></p></li>\n</ul>\n\n<p>FYI: This scripting style is borrowed from this:\n<a href=\"https://github.com/lopuhin/kaggle-script-template\">https://github.com/lopuhin/kaggle-script-template</a></p>\n\n<p>I prefer script style, which is easy to manage by git.</p>\n\n<h1>summary</h1>\n\n<p>To be honest, there is no special technique in my solution.\nI chose stylegan, then looked for good parameters.\nMy stylegan starts with public score 140, then tuning parameters, I got 35.</p>\n\n<ul>\n<li>for preprocessing: crop image by bounding box, random horizontal flip</li>\n<li>reduce input channel size of style generator to 256 because of resource limitation.</li>\n</ul>\n\n<p>all hyper parameters are shown in below.\n<a href=\"https://github.com/katotetsuro/chainer-stylegan/blob/ac/src/stylegan/config.py\">https://github.com/katotetsuro/chainer-stylegan/blob/ac/src/stylegan/config.py</a>\n<a href=\"https://github.com/katotetsuro/chainer-stylegan/blob/ac/script_template.py\">https://github.com/katotetsuro/chainer-stylegan/blob/ac/script_template.py</a></p>\n\n<h1>trials that didn't work for me</h1>\n\n<ul>\n<li>Truncation trick from stylegan paper. I tried various truncation strength, but without truncation was the best.</li>\n<li>Auxiliary Classifier</li>\n<li>Data cleaning. I thought training data quality is quite important, so I did some procedure (dog face detection, manually picking good images..) but nothing worked.\nToday, I read some kernels and learned  there's some good techniques.\nThis is one of the room for improvement of my solution, I guess.</li>\n</ul>\n\n<p>It was really fun competition! Thanks for kaggle team!!</p>",
      "rawMarkdown": "# code\n\n- kernel\nhttps://www.kaggle.com/katotetsuro/stylegan-private-89?scriptVersionId=18524688\n\n- github\nhttps://github.com/katotetsuro/chainer-stylegan/tree/ac\n\nFYI: This scripting style is borrowed from this:\nhttps://github.com/lopuhin/kaggle-script-template\n\nI prefer script style, which is easy to manage by git.\n\n# summary\n\nTo be honest, there is no special technique in my solution.\nI chose stylegan, then looked for good parameters.\nMy stylegan starts with public score 140, then tuning parameters, I got 35.\n\n- for preprocessing: crop image by bounding box, random horizontal flip\n- reduce input channel size of style generator to 256 because of resource limitation.\n\nall hyper parameters are shown in below.\nhttps://github.com/katotetsuro/chainer-stylegan/blob/ac/src/stylegan/config.py\nhttps://github.com/katotetsuro/chainer-stylegan/blob/ac/script_template.py\n\n# trials that didn't work for me\n\n- Truncation trick from stylegan paper. I tried various truncation strength, but without truncation was the best.\n- Auxiliary Classifier\n- Data cleaning. I thought training data quality is quite important, so I did some procedure (dog face detection, manually picking good images..) but nothing worked.\nToday, I read some kernels and learned  there's some good techniques.\nThis is one of the room for improvement of my solution, I guess.\n\nIt was really fun competition! Thanks for kaggle team!!",
      "votes": null
    },
    {
      "id": "612460",
      "postDate": "08/29/2019 19:57:00",
      "content": "<p>Great job tkato!</p>",
      "rawMarkdown": "Great job tkato!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 612460,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "08/29/2019 19:57:00",
      "content": "<p>Great job tkato!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "612005": "# code\n\n- kernel\nhttps://www.kaggle.com/katotetsuro/stylegan-private-89?scriptVersionId=18524688\n\n- github\nhttps://github.com/katotetsuro/chainer-stylegan/tree/ac\n\nFYI: This scripting style is borrowed from this:\nhttps://github.com/lopuhin/kaggle-script-template\n\nI prefer script style, which is easy to manage by git.\n\n# summary\n\nTo be honest, there is no special technique in my solution.\nI chose stylegan, then looked for good parameters.\nMy stylegan starts with public score 140, then tuning parameters, I got 35.\n\n- for preprocessing: crop image by bounding box, random horizontal flip\n- reduce input channel size of style generator to 256 because of resource limitation.\n\nall hyper parameters are shown in below.\nhttps://github.com/katotetsuro/chainer-stylegan/blob/ac/src/stylegan/config.py\nhttps://github.com/katotetsuro/chainer-stylegan/blob/ac/script_template.py\n\n# trials that didn't work for me\n\n- Truncation trick from stylegan paper. I tried various truncation strength, but without truncation was the best.\n- Auxiliary Classifier\n- Data cleaning. I thought training data quality is quite important, so I did some procedure (dog face detection, manually picking good images..) but nothing worked.\nToday, I read some kernels and learned  there's some good techniques.\nThis is one of the room for improvement of my solution, I guess.\n\nIt was really fun competition! Thanks for kaggle team!!",
    "612460": "Great job tkato!"
  },
  "source": "existing_discussion",
  "complete": true,
  "competition_id": 15062,
  "competition_slug": "generative-dog-images",
  "team_id": 3345987,
  "solution_url": "/competitions/generative-dog-images/writeups/tkato-8th-place-solution-private-89"
}