{
  "id": 102100,
  "title": "A simple way to determine if the method is a Memorized GAN",
  "url": "/competitions/generative-dog-images/discussion/102100",
  "author_name": "",
  "post_date": "2019-07-31T01:03:16.105904300Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>In my understanding of method Memorized GAN，it simply memorizes the required number of images from the training data set. When it generates a picture, it has to make sure that the input vector is in the training format, which means it's not completely 'generated'.\n<strong>So i think it's easy to ban it</strong>\n*<em>When generate submit images,check whether the input vector to the Generator Network is truly random or fake random(just specific vectors)</em>*</p>",
  "messages": [
    {
      "id": "588731",
      "postDate": "07/31/2019 01:03:16",
      "content": "<p>In my understanding of method Memorized GAN，it simply memorizes the required number of images from the training data set. When it generates a picture, it has to make sure that the input vector is in the training format, which means it's not completely 'generated'.\n<strong>So i think it's easy to ban it</strong>\n*<em>When generate submit images,check whether the input vector to the Generator Network is truly random or fake random(just specific vectors)</em>*</p>",
      "rawMarkdown": "In my understanding of method Memorized GAN，it simply memorizes the required number of images from the training data set. When it generates a picture, it has to make sure that the input vector is in the training format, which means it's not completely 'generated'.\n**So i think it's easy to ban it**\n**When generate submit images,check whether the input vector to the Generator Network is truly random or fake random(just specific vectors)**",
      "votes": null
    },
    {
      "id": "588835",
      "postDate": "07/31/2019 05:41:37",
      "content": "<p>They've already been allowed by the kaggle team. I don't think it's possible to explicitly ban memorizer GANs.\nAlthough, I think kaggle must've made sure memorizer GANs perform terribly on the private dataset and model.</p>",
      "rawMarkdown": "They've already been allowed by the kaggle team. I don't think it's possible to explicitly ban memorizer GANs.\nAlthough, I think kaggle must've made sure memorizer GANs perform terribly on the private dataset and model.",
      "votes": null
    },
    {
      "id": "590689",
      "postDate": "08/02/2019 13:20:57",
      "content": "<p>Checking input vectors won't distinguish memorizer GANs. You can make a neural network take a random vector of length 100, then have a dense layer of 10000, then have a softmax layer. You can train this front-end module to map random vectors to specific vectors and continue your generator from there.</p>\n\n<p>Then based on probability if you input 10000 random vectors, you're expected to get 6000 out of 10000 unique softmax outputs. That is enough unique images to score well and beat any traditional GAN.</p>",
      "rawMarkdown": "Checking input vectors won't distinguish memorizer GANs. You can make a neural network take a random vector of length 100, then have a dense layer of 10000, then have a softmax layer. You can train this front-end module to map random vectors to specific vectors and continue your generator from there.\n\nThen based on probability if you input 10000 random vectors, you're expected to get 6000 out of 10000 unique softmax outputs. That is enough unique images to score well and beat any traditional GAN.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 588835,
      "author_name": "timetraveller98",
      "author_url": "",
      "post_date": "07/31/2019 05:41:37",
      "content": "<p>They've already been allowed by the kaggle team. I don't think it's possible to explicitly ban memorizer GANs.\nAlthough, I think kaggle must've made sure memorizer GANs perform terribly on the private dataset and model.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 590689,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "08/02/2019 13:20:57",
      "content": "<p>Checking input vectors won't distinguish memorizer GANs. You can make a neural network take a random vector of length 100, then have a dense layer of 10000, then have a softmax layer. You can train this front-end module to map random vectors to specific vectors and continue your generator from there.</p>\n\n<p>Then based on probability if you input 10000 random vectors, you're expected to get 6000 out of 10000 unique softmax outputs. That is enough unique images to score well and beat any traditional GAN.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "588731": "In my understanding of method Memorized GAN，it simply memorizes the required number of images from the training data set. When it generates a picture, it has to make sure that the input vector is in the training format, which means it's not completely 'generated'.\n**So i think it's easy to ban it**\n**When generate submit images,check whether the input vector to the Generator Network is truly random or fake random(just specific vectors)**",
    "588835": "They've already been allowed by the kaggle team. I don't think it's possible to explicitly ban memorizer GANs.\nAlthough, I think kaggle must've made sure memorizer GANs perform terribly on the private dataset and model.",
    "590689": "Checking input vectors won't distinguish memorizer GANs. You can make a neural network take a random vector of length 100, then have a dense layer of 10000, then have a softmax layer. You can train this front-end module to map random vectors to specific vectors and continue your generator from there.\n\nThen based on probability if you input 10000 random vectors, you're expected to get 6000 out of 10000 unique softmax outputs. That is enough unique images to score well and beat any traditional GAN."
  },
  "source": "meta"
}