{
  "id": 99938,
  "title": "Strange Discriminator Loss Surge",
  "url": "/competitions/generative-dog-images/discussion/99938",
  "author_name": "",
  "post_date": "2019-07-15T13:03:59.224275200Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F114413%2F05fbb22610fe0e3baf20dcd453f7e3aa%2F__results___21_1.png?generation=1563195757337206&amp;alt=media\" alt=\"\">\nThe blue is discriminator loss. It have a sudden increase in discriminator loss and all the image after that become black. What actually happen?</p>",
  "messages": [
    {
      "id": "575464",
      "postDate": "07/15/2019 13:03:59",
      "content": "<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F114413%2F05fbb22610fe0e3baf20dcd453f7e3aa%2F__results___21_1.png?generation=1563195757337206&amp;alt=media\" alt=\"\">\nThe blue is discriminator loss. It have a sudden increase in discriminator loss and all the image after that become black. What actually happen?</p>",
      "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F114413%2F05fbb22610fe0e3baf20dcd453f7e3aa%2F__results___21_1.png?generation=1563195757337206&amp;alt=media)\nThe blue is discriminator loss. It have a sudden increase in discriminator loss and all the image after that become black. What actually happen?",
      "votes": null
    },
    {
      "id": "575556",
      "postDate": "07/15/2019 16:23:26",
      "content": "<p>Whicj type of preprocessing are you using? Which type of optimizer and lr are you using? </p>",
      "rawMarkdown": "Whicj type of preprocessing are you using? Which type of optimizer and lr are you using?",
      "votes": null
    },
    {
      "id": "576862",
      "postDate": "07/16/2019 05:35:22",
      "content": "<p>wgan, preprocess image to [-1,1] Adam, lr = 0.0002, beta1 = 0.5</p>",
      "rawMarkdown": "wgan, preprocess image to [-1,1] Adam, lr = 0.0002, beta1 = 0.5",
      "votes": null
    },
    {
      "id": "577826",
      "postDate": "07/17/2019 05:58:58",
      "content": "<p>Maybe use <a href=\"https://arxiv.org/pdf/1704.00028.pdf\">WGAN-GP</a> instead can solve this situation?\nTo switch WGAN to WGAN-GP you need two steps:\n1. Remove weight clipping\n2. Add gradient penalty term to your loss function</p>",
      "rawMarkdown": "Maybe use [WGAN-GP](https://arxiv.org/pdf/1704.00028.pdf) instead can solve this situation?\nTo switch WGAN to WGAN-GP you need two steps:\n1. Remove weight clipping\n2. Add gradient penalty term to your loss function",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 575556,
      "author_name": "enric1296",
      "author_url": "",
      "post_date": "07/15/2019 16:23:26",
      "content": "<p>Whicj type of preprocessing are you using? Which type of optimizer and lr are you using? </p>",
      "votes": null,
      "replies": [
        {
          "id": 576862,
          "author_name": "skyap79",
          "author_url": "",
          "post_date": "07/16/2019 05:35:22",
          "content": "<p>wgan, preprocess image to [-1,1] Adam, lr = 0.0002, beta1 = 0.5</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 577826,
      "author_name": "yanjenchen",
      "author_url": "",
      "post_date": "07/17/2019 05:58:58",
      "content": "<p>Maybe use <a href=\"https://arxiv.org/pdf/1704.00028.pdf\">WGAN-GP</a> instead can solve this situation?\nTo switch WGAN to WGAN-GP you need two steps:\n1. Remove weight clipping\n2. Add gradient penalty term to your loss function</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "575464": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F114413%2F05fbb22610fe0e3baf20dcd453f7e3aa%2F__results___21_1.png?generation=1563195757337206&amp;alt=media)\nThe blue is discriminator loss. It have a sudden increase in discriminator loss and all the image after that become black. What actually happen?",
    "575556": "Whicj type of preprocessing are you using? Which type of optimizer and lr are you using?",
    "576862": "wgan, preprocess image to [-1,1] Adam, lr = 0.0002, beta1 = 0.5",
    "577826": "Maybe use [WGAN-GP](https://arxiv.org/pdf/1704.00028.pdf) instead can solve this situation?\nTo switch WGAN to WGAN-GP you need two steps:\n1. Remove weight clipping\n2. Add gradient penalty term to your loss function"
  },
  "source": "meta"
}