{
  "id": 101494,
  "title": "A question about \"memorization\" GANs",
  "url": "/competitions/generative-dog-images/discussion/101494",
  "author_name": "",
  "post_date": "2019-07-26T08:58:18.286088800Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I've seen a several kernels where the authors implement \"memorizer\" networks. However, I still not sure how they really work :) </p>\n\n<p>Do these networks just interpolate between images of the original dataset? If so then this approach will not generate  \"brand-new\" images, right? Like, it will always combine, or blend the samples from the training set, probably in some tricky way but still. Does it mean that these models heavily overfit the data and don't generalize?</p>\n\n<p>I would like to find out a GAN model/architecture capable to generate new images, and as \"original\" and \"realistic\" as possible. I understand that it could take a lot of time to train such a model but that's a second question. So the memorization networks are not a way to go to achieve this goal, right?</p>",
  "messages": [
    {
      "id": "584649",
      "postDate": "07/26/2019 08:58:18",
      "content": "<p>I've seen a several kernels where the authors implement \"memorizer\" networks. However, I still not sure how they really work :) </p>\n\n<p>Do these networks just interpolate between images of the original dataset? If so then this approach will not generate  \"brand-new\" images, right? Like, it will always combine, or blend the samples from the training set, probably in some tricky way but still. Does it mean that these models heavily overfit the data and don't generalize?</p>\n\n<p>I would like to find out a GAN model/architecture capable to generate new images, and as \"original\" and \"realistic\" as possible. I understand that it could take a lot of time to train such a model but that's a second question. So the memorization networks are not a way to go to achieve this goal, right?</p>",
      "rawMarkdown": "I've seen a several kernels where the authors implement \"memorizer\" networks. However, I still not sure how they really work :) \n\nDo these networks just interpolate between images of the original dataset? If so then this approach will not generate  \"brand-new\" images, right? Like, it will always combine, or blend the samples from the training set, probably in some tricky way but still. Does it mean that these models heavily overfit the data and don't generalize?\n\nI would like to find out a GAN model/architecture capable to generate new images, and as \"original\" and \"realistic\" as possible. I understand that it could take a lot of time to train such a model but that's a second question. So the memorization networks are not a way to go to achieve this goal, right?",
      "votes": null
    },
    {
      "id": "584653",
      "postDate": "07/26/2019 09:08:39",
      "content": "<p>Yes. They are garbage and serve no purpose. \nIt's just a way to recreate the original images, as you said, it's just heavily overfitted on the input data. It creates nothing new and doesn't know how to generalize at all.\nFollow this post by Chris: <a href=\"https://www.kaggle.com/c/generative-dog-images/discussion/99215\">Memorizer GANs</a>. It will give you great insights on how memorizer GAN kernels work.</p>",
      "rawMarkdown": "Yes. They are garbage and serve no purpose. \nIt's just a way to recreate the original images, as you said, it's just heavily overfitted on the input data. It creates nothing new and doesn't know how to generalize at all.\nFollow this post by Chris: [Memorizer GANs](https://www.kaggle.com/c/generative-dog-images/discussion/99215). It will give you great insights on how memorizer GAN kernels work.",
      "votes": null
    },
    {
      "id": "591189",
      "postDate": "08/03/2019 09:34:20",
      "content": "<p>Ok great, thank you for a link!</p>",
      "rawMarkdown": "Ok great, thank you for a link!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 584653,
      "author_name": "timetraveller98",
      "author_url": "",
      "post_date": "07/26/2019 09:08:39",
      "content": "<p>Yes. They are garbage and serve no purpose. \nIt's just a way to recreate the original images, as you said, it's just heavily overfitted on the input data. It creates nothing new and doesn't know how to generalize at all.\nFollow this post by Chris: <a href=\"https://www.kaggle.com/c/generative-dog-images/discussion/99215\">Memorizer GANs</a>. It will give you great insights on how memorizer GAN kernels work.</p>",
      "votes": null,
      "replies": [
        {
          "id": 591189,
          "author_name": "purplejester",
          "author_url": "",
          "post_date": "08/03/2019 09:34:20",
          "content": "<p>Ok great, thank you for a link!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "584649": "I've seen a several kernels where the authors implement \"memorizer\" networks. However, I still not sure how they really work :) \n\nDo these networks just interpolate between images of the original dataset? If so then this approach will not generate  \"brand-new\" images, right? Like, it will always combine, or blend the samples from the training set, probably in some tricky way but still. Does it mean that these models heavily overfit the data and don't generalize?\n\nI would like to find out a GAN model/architecture capable to generate new images, and as \"original\" and \"realistic\" as possible. I understand that it could take a lot of time to train such a model but that's a second question. So the memorization networks are not a way to go to achieve this goal, right?",
    "584653": "Yes. They are garbage and serve no purpose. \nIt's just a way to recreate the original images, as you said, it's just heavily overfitted on the input data. It creates nothing new and doesn't know how to generalize at all.\nFollow this post by Chris: [Memorizer GANs](https://www.kaggle.com/c/generative-dog-images/discussion/99215). It will give you great insights on how memorizer GAN kernels work.",
    "591189": "Ok great, thank you for a link!"
  },
  "source": "meta"
}