{
  "id": 98504,
  "title": "StyleGAN",
  "url": "/competitions/generative-dog-images/discussion/98504",
  "author_name": "hirune924",
  "post_date": "2019-07-04T08:22:02.409000",
  "votes": 4,
  "comment_count": 11,
  "views": 0,
  "content": "<p>I want to create very clear dogs.\nSo I trying to train StyleGAN.  but it's very difficult because of kernel time limit...</p>",
  "messages": [
    {
      "id": 567987,
      "postDate": "2019-07-04T08:22:02.410Z",
      "content": "<p>I want to create very clear dogs.\nSo I trying to train StyleGAN.  but it's very difficult because of kernel time limit...</p>",
      "rawMarkdown": "I want to create very clear dogs.\nSo I trying to train StyleGAN.  but it's very difficult because of kernel time limit...",
      "votes": 4
    },
    {
      "id": 568540,
      "postDate": "2019-07-05T04:47:05.760Z",
      "content": "<p>You could try ProGAN. </p>",
      "rawMarkdown": "You could try ProGAN. ",
      "votes": 1
    },
    {
      "id": 568130,
      "postDate": "2019-07-04T13:01:58.020Z",
      "content": "<p>So GANs have a serious lack of having a true measuring metric. The metric (FID) used in this contest although looks seemingly promising. Style GAN are certainly not the best choice for this contest as they take really long time to train. You can perhaps try to WGAN with a number of tweaks.  The default hyper-parameters mentioned in its paper also works fine.</p>",
      "rawMarkdown": "So GANs have a serious lack of having a true measuring metric. The metric (FID) used in this contest although looks seemingly promising. Style GAN are certainly not the best choice for this contest as they take really long time to train. You can perhaps try to WGAN with a number of tweaks.  The default hyper-parameters mentioned in its paper also works fine.\n",
      "votes": 1,
      "replies": [
        {
          "id": 568146,
          "postDate": "2019-07-04T13:27:50.203Z",
          "content": "<p>I have tried with WGAN . I would not say it is promising or whether I would really get a puppy as output . But it probably is faster than StyleGAN </p>\n\n<p>My kernel : \n<a href=\"https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation?scriptVersionId=16700224\">https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation?scriptVersionId=16700224</a></p>",
          "rawMarkdown": "I have tried with WGAN . I would not say it is promising or whether I would really get a puppy as output . But it probably is faster than StyleGAN \n\nMy kernel : \nhttps://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation?scriptVersionId=16700224",
          "votes": 1
        },
        {
          "id": 568150,
          "postDate": "2019-07-04T13:36:08.887Z",
          "content": "<p>Good work <a href=\"/phoenix9032\">@phoenix9032</a>, I am facing a similar problem of <code>memory allocation fail</code> implementing fastai notebooks. 🤒 Do let me know how you solve this.</p>",
          "rawMarkdown": "Good work @phoenix9032, I am facing a similar problem of `memory allocation fail` implementing fastai notebooks. 🤒 Do let me know how you solve this."
        },
        {
          "id": 568154,
          "postDate": "2019-07-04T13:40:25.583Z",
          "content": "<p>Nope .. I thought I executed multiple kernel commits together .. that's why . If that's not the problem I don't know what can cause this . I resubmitted with 200 epochs again .Will keep my fingers crossed . My Kernel  version 6 is good . I would expect Kernel to fail the commit in case of memory issue so that people can view the right Kernel . That's a bummer :( </p>",
          "rawMarkdown": "Nope .. I thought I executed multiple kernel commits together .. that's why . If that's not the problem I don't know what can cause this . I resubmitted with 200 epochs again .Will keep my fingers crossed . My Kernel  version 6 is good . I would expect Kernel to fail the commit in case of memory issue so that people can view the right Kernel . That's a bummer :( "
        }
      ]
    },
    {
      "id": 568881,
      "postDate": "2019-07-05T15:53:41.897Z",
      "content": "<p>There are good GAN zoo.\n<a href=\"https://github.com/hindupuravinash/the-gan-zoo\">https://github.com/hindupuravinash/the-gan-zoo</a></p>",
      "rawMarkdown": "There are good GAN zoo.\nhttps://github.com/hindupuravinash/the-gan-zoo"
    },
    {
      "id": 568122,
      "postDate": "2019-07-04T12:41:48.933Z",
      "content": "<p>That's the treadoff having a free platform to try out new things and learn .</p>\n\n<p>Can Stylegan give good result here ? I thought that is for different kind of problem . Any idea ?</p>",
      "rawMarkdown": "That's the treadoff having a free platform to try out new things and learn .\n\nCan Stylegan give good result here ? I thought that is for different kind of problem . Any idea ?",
      "replies": [
        {
          "id": 568159,
          "postDate": "2019-07-04T13:55:29.073Z",
          "content": "<p>I already tried some speedup method(small network or small batch size or something...) on training StyleGAN, but it's not enough...\nIf GAN competition will be held again, i want more long kernel time. \n(And I just want to create good dogs regardless of winning or losing)</p>",
          "rawMarkdown": "I already tried some speedup method(small network or small batch size or something...) on training StyleGAN, but it's not enough...\nIf GAN competition will be held again, i want more long kernel time. \n(And I just want to create good dogs regardless of winning or losing)",
          "votes": 4
        },
        {
          "id": 568472,
          "postDate": "2019-07-05T01:38:55.013Z",
          "content": "<p>\b<a href=\"/hirune924\">@hirune924</a> How long did it take to run your StyleGAN with speedup method? </p>",
          "rawMarkdown": "\b@hirune924 How long did it take to run your StyleGAN with speedup method? "
        },
        {
          "id": 568477,
          "postDate": "2019-07-05T01:51:22.487Z",
          "content": "<p>I run about 140 epoch for training on kernel limit. \nDue to the nature of StyleGAN, the calculation gets slower gradually.</p>",
          "rawMarkdown": "I run about 140 epoch for training on kernel limit. \nDue to the nature of StyleGAN, the calculation gets slower gradually.",
          "votes": 1
        }
      ]
    },
    {
      "id": 568845,
      "postDate": "2019-07-05T14:34:06.650Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 568540,
      "author_name": "sh",
      "author_url": "",
      "post_date": "2019-07-05T04:47:05.760000",
      "content": "<p>You could try ProGAN. </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 568130,
      "author_name": "Tushar",
      "author_url": "",
      "post_date": "2019-07-04T13:01:58.020000",
      "content": "<p>So GANs have a serious lack of having a true measuring metric. The metric (FID) used in this contest although looks seemingly promising. Style GAN are certainly not the best choice for this contest as they take really long time to train. You can perhaps try to WGAN with a number of tweaks.  The default hyper-parameters mentioned in its paper also works fine.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 568146,
          "author_name": "Nirjhar Roy",
          "author_url": "",
          "post_date": "2019-07-04T13:27:50.203000",
          "content": "<p>I have tried with WGAN . I would not say it is promising or whether I would really get a puppy as output . But it probably is faster than StyleGAN </p>\n\n<p>My kernel : \n<a href=\"https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation?scriptVersionId=16700224\">https://www.kaggle.com/phoenix9032/fastai-wgan-for-dog-image-generation?scriptVersionId=16700224</a></p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 568150,
          "author_name": "Tushar",
          "author_url": "",
          "post_date": "2019-07-04T13:36:08.887000",
          "content": "<p>Good work <a href=\"/phoenix9032\">@phoenix9032</a>, I am facing a similar problem of <code>memory allocation fail</code> implementing fastai notebooks. 🤒 Do let me know how you solve this.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 568154,
          "author_name": "Nirjhar Roy",
          "author_url": "",
          "post_date": "2019-07-04T13:40:25.583000",
          "content": "<p>Nope .. I thought I executed multiple kernel commits together .. that's why . If that's not the problem I don't know what can cause this . I resubmitted with 200 epochs again .Will keep my fingers crossed . My Kernel  version 6 is good . I would expect Kernel to fail the commit in case of memory issue so that people can view the right Kernel . That's a bummer :( </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 568881,
      "author_name": "makogarei",
      "author_url": "",
      "post_date": "2019-07-05T15:53:41.897000",
      "content": "<p>There are good GAN zoo.\n<a href=\"https://github.com/hindupuravinash/the-gan-zoo\">https://github.com/hindupuravinash/the-gan-zoo</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 568122,
      "author_name": "Nirjhar Roy",
      "author_url": "",
      "post_date": "2019-07-04T12:41:48.933000",
      "content": "<p>That's the treadoff having a free platform to try out new things and learn .</p>\n\n<p>Can Stylegan give good result here ? I thought that is for different kind of problem . Any idea ?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 568159,
          "author_name": "hirune924",
          "author_url": "",
          "post_date": "2019-07-04T13:55:29.073000",
          "content": "<p>I already tried some speedup method(small network or small batch size or something...) on training StyleGAN, but it's not enough...\nIf GAN competition will be held again, i want more long kernel time. \n(And I just want to create good dogs regardless of winning or losing)</p>",
          "votes": 4,
          "replies": []
        },
        {
          "id": 568472,
          "author_name": "Pondel",
          "author_url": "",
          "post_date": "2019-07-05T01:38:55.013000",
          "content": "<p>\b<a href=\"/hirune924\">@hirune924</a> How long did it take to run your StyleGAN with speedup method? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 568477,
          "author_name": "hirune924",
          "author_url": "",
          "post_date": "2019-07-05T01:51:22.487000",
          "content": "<p>I run about 140 epoch for training on kernel limit. \nDue to the nature of StyleGAN, the calculation gets slower gradually.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 568845,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-07-05T14:34:06.650000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "567987": "I want to create very clear dogs.\nSo I trying to train StyleGAN.  but it's very difficult because of kernel time limit...",
    "568540": "You could try ProGAN. ",
    "568130": "So GANs have a serious lack of having a true measuring metric. The metric (FID) used in this contest although looks seemingly promising. Style GAN are certainly not the best choice for this contest as they take really long time to train. You can perhaps try to WGAN with a number of tweaks.  The default hyper-parameters mentioned in its paper also works fine.\n",
    "568881": "There are good GAN zoo.\nhttps://github.com/hindupuravinash/the-gan-zoo",
    "568122": "That's the treadoff having a free platform to try out new things and learn .\n\nCan Stylegan give good result here ? I thought that is for different kind of problem . Any idea ?",
    "568845": ""
  }
}