{
  "id": 98347,
  "title": "Running DCGAN on cropped images",
  "url": "/competitions/generative-dog-images/discussion/98347",
  "author_name": "",
  "post_date": "2019-07-03T05:37:40.835489400Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I though cropping the image will do great improvement over normal noisy data. But I think It still producing the similar data. </p>\n\n<p>Have a look at this kernel : <a href=\"https://www.kaggle.com/whizzkid/dcgan-on-cropped-dog-images\">https://www.kaggle.com/whizzkid/dcgan-on-cropped-dog-images</a></p>\n\n<p>Does that means GANs picks up right information even from noisy environment ?</p>",
  "messages": [
    {
      "id": "567149",
      "postDate": "07/03/2019 05:37:40",
      "content": "<p>I though cropping the image will do great improvement over normal noisy data. But I think It still producing the similar data. </p>\n\n<p>Have a look at this kernel : <a href=\"https://www.kaggle.com/whizzkid/dcgan-on-cropped-dog-images\">https://www.kaggle.com/whizzkid/dcgan-on-cropped-dog-images</a></p>\n\n<p>Does that means GANs picks up right information even from noisy environment ?</p>",
      "rawMarkdown": "I though cropping the image will do great improvement over normal noisy data. But I think It still producing the similar data. \n\nHave a look at this kernel : https://www.kaggle.com/whizzkid/dcgan-on-cropped-dog-images\n\nDoes that means GANs picks up right information even from noisy environment ?",
      "votes": null
    },
    {
      "id": "567264",
      "postDate": "07/03/2019 08:53:48",
      "content": "<p>Sure it will mostly pick up the dogs, as they are the only thing that is present in ALL pictures. \nBut recent papers showed that GANs are much easier to train and to converge on picture classes that are similar. (e.g. faces instead of full bodys mixed with faces). \nRead: specialized GANs are easier than multipurpose GANs</p>",
      "rawMarkdown": "Sure it will mostly pick up the dogs, as they are the only thing that is present in ALL pictures. \nBut recent papers showed that GANs are much easier to train and to converge on picture classes that are similar. (e.g. faces instead of full bodys mixed with faces). \nRead: specialized GANs are easier than multipurpose GANs",
      "votes": null
    },
    {
      "id": "567302",
      "postDate": "07/03/2019 10:17:34",
      "content": "<p>Could you please provide the link. <a href=\"/ilu000\">@ilu000</a> </p>",
      "rawMarkdown": "Could you please provide the link. @ilu000",
      "votes": null
    },
    {
      "id": "567310",
      "postDate": "07/03/2019 10:25:36",
      "content": "<p>Sure, <a href=\"https://arxiv.org/pdf/1710.10196.pdf\">https://arxiv.org/pdf/1710.10196.pdf</a> APPENDIX C</p>",
      "rawMarkdown": "Sure, https://arxiv.org/pdf/1710.10196.pdf APPENDIX C",
      "votes": null
    },
    {
      "id": "567322",
      "postDate": "07/03/2019 10:51:58",
      "content": "<p>Thanks for the link. This makes sense. CNNs (and therefore GANs) learn by finding patterns (and patterns of patterns) within images. If your training images have too many patterns (too much image variety), then your CNN (GAN) cannot specialize in learning a smaller set of patterns. </p>\n\n<p>Furthermore to learn patterns you needs lots of training examples. Using 20000 dogs is a higher ratio of training images to patterns, then 20000 images of landscapes that include dogs.</p>",
      "rawMarkdown": "Thanks for the link. This makes sense. CNNs (and therefore GANs) learn by finding patterns (and patterns of patterns) within images. If your training images have too many patterns (too much image variety), then your CNN (GAN) cannot specialize in learning a smaller set of patterns. \n\nFurthermore to learn patterns you needs lots of training examples. Using 20000 dogs is a higher ratio of training images to patterns, then 20000 images of landscapes that include dogs.",
      "votes": null
    },
    {
      "id": "567358",
      "postDate": "07/03/2019 11:48:17",
      "content": "<p>Especially this part is very interesting:\n<code>\nWe perform the above processing for all 202599 images in the dataset, analyze the resulting 1024 × 1024 images further to estimate the final image quality, sort the images accordingly, and discard all but the best 30000 images. We use a frequency-based quality metric that favors images whose power spectrum contains a broad range of frequencies and is approximately radially symmetric. This penalizes blurry images as well as images that have conspicuous directional features due to, e.g., visible halftoning patterns. We selected the cutoff point of 30000 images as a practical sweet spot between variation and image quality, because it appeared to yield the best results.\n</code></p>",
      "rawMarkdown": "Especially this part is very interesting:\n```\nWe perform the above processing for all 202599 images in the dataset, analyze the resulting 1024 × 1024 images further to estimate the final image quality, sort the images accordingly, and discard all but the best 30000 images. We use a frequency-based quality metric that favors images whose power spectrum contains a broad range of frequencies and is approximately radially symmetric. This penalizes blurry images as well as images that have conspicuous directional features due to, e.g., visible halftoning patterns. We selected the cutoff point of 30000 images as a practical sweet spot between variation and image quality, because it appeared to yield the best results.\n```",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 567264,
      "author_name": "ilu000",
      "author_url": "",
      "post_date": "07/03/2019 08:53:48",
      "content": "<p>Sure it will mostly pick up the dogs, as they are the only thing that is present in ALL pictures. \nBut recent papers showed that GANs are much easier to train and to converge on picture classes that are similar. (e.g. faces instead of full bodys mixed with faces). \nRead: specialized GANs are easier than multipurpose GANs</p>",
      "votes": null,
      "replies": [
        {
          "id": 567302,
          "author_name": "shwetagoyal4",
          "author_url": "",
          "post_date": "07/03/2019 10:17:34",
          "content": "<p>Could you please provide the link. <a href=\"/ilu000\">@ilu000</a> </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 567310,
          "author_name": "ilu000",
          "author_url": "",
          "post_date": "07/03/2019 10:25:36",
          "content": "<p>Sure, <a href=\"https://arxiv.org/pdf/1710.10196.pdf\">https://arxiv.org/pdf/1710.10196.pdf</a> APPENDIX C</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 567322,
          "author_name": "cdeotte",
          "author_url": "",
          "post_date": "07/03/2019 10:51:58",
          "content": "<p>Thanks for the link. This makes sense. CNNs (and therefore GANs) learn by finding patterns (and patterns of patterns) within images. If your training images have too many patterns (too much image variety), then your CNN (GAN) cannot specialize in learning a smaller set of patterns. </p>\n\n<p>Furthermore to learn patterns you needs lots of training examples. Using 20000 dogs is a higher ratio of training images to patterns, then 20000 images of landscapes that include dogs.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 567358,
          "author_name": "ilu000",
          "author_url": "",
          "post_date": "07/03/2019 11:48:17",
          "content": "<p>Especially this part is very interesting:\n<code>\nWe perform the above processing for all 202599 images in the dataset, analyze the resulting 1024 × 1024 images further to estimate the final image quality, sort the images accordingly, and discard all but the best 30000 images. We use a frequency-based quality metric that favors images whose power spectrum contains a broad range of frequencies and is approximately radially symmetric. This penalizes blurry images as well as images that have conspicuous directional features due to, e.g., visible halftoning patterns. We selected the cutoff point of 30000 images as a practical sweet spot between variation and image quality, because it appeared to yield the best results.\n</code></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "567149": "I though cropping the image will do great improvement over normal noisy data. But I think It still producing the similar data. \n\nHave a look at this kernel : https://www.kaggle.com/whizzkid/dcgan-on-cropped-dog-images\n\nDoes that means GANs picks up right information even from noisy environment ?",
    "567264": "Sure it will mostly pick up the dogs, as they are the only thing that is present in ALL pictures. \nBut recent papers showed that GANs are much easier to train and to converge on picture classes that are similar. (e.g. faces instead of full bodys mixed with faces). \nRead: specialized GANs are easier than multipurpose GANs",
    "567302": "Could you please provide the link. @ilu000",
    "567310": "Sure, https://arxiv.org/pdf/1710.10196.pdf APPENDIX C",
    "567322": "Thanks for the link. This makes sense. CNNs (and therefore GANs) learn by finding patterns (and patterns of patterns) within images. If your training images have too many patterns (too much image variety), then your CNN (GAN) cannot specialize in learning a smaller set of patterns. \n\nFurthermore to learn patterns you needs lots of training examples. Using 20000 dogs is a higher ratio of training images to patterns, then 20000 images of landscapes that include dogs.",
    "567358": "Especially this part is very interesting:\n```\nWe perform the above processing for all 202599 images in the dataset, analyze the resulting 1024 × 1024 images further to estimate the final image quality, sort the images accordingly, and discard all but the best 30000 images. We use a frequency-based quality metric that favors images whose power spectrum contains a broad range of frequencies and is approximately radially symmetric. This penalizes blurry images as well as images that have conspicuous directional features due to, e.g., visible halftoning patterns. We selected the cutoff point of 30000 images as a practical sweet spot between variation and image quality, because it appeared to yield the best results.\n```"
  },
  "source": "meta"
}