{
  "id": 102596,
  "title": "which augmentation method helps for gan??",
  "url": "/competitions/generative-dog-images/discussion/102596",
  "author_name": "",
  "post_date": "2019-08-03T04:05:29.983736200Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I haven't tried many augmentations, but I want to know what others think about the augmenations.</p>\n\n<p>I've tried with fixed augmentation as resize and centercrop (64 size)</p>\n\n<p>and used horizontal flip + random rotation as baseline</p>\n\n<p>how are you guys dealing with this problem?</p>\n\n<p>are the augmentations really helpful for training GANs?</p>\n\n<p>I would really appreciate it if you share your experiments, thanks!</p>",
  "messages": [
    {
      "id": "591050",
      "postDate": "08/03/2019 04:05:29",
      "content": "<p>I haven't tried many augmentations, but I want to know what others think about the augmenations.</p>\n\n<p>I've tried with fixed augmentation as resize and centercrop (64 size)</p>\n\n<p>and used horizontal flip + random rotation as baseline</p>\n\n<p>how are you guys dealing with this problem?</p>\n\n<p>are the augmentations really helpful for training GANs?</p>\n\n<p>I would really appreciate it if you share your experiments, thanks!</p>",
      "rawMarkdown": "I haven't tried many augmentations, but I want to know what others think about the augmenations.\n\nI've tried with fixed augmentation as resize and centercrop (64 size)\n\nand used horizontal flip + random rotation as baseline\n\nhow are you guys dealing with this problem?\n\nare the augmentations really helpful for training GANs?\n\nI would really appreciate it if you share your experiments, thanks!",
      "votes": null
    },
    {
      "id": "591230",
      "postDate": "08/03/2019 11:01:59",
      "content": "<p>I also would be glad to know more. The only thing I am using right now is horizontal flip and cropping. I tried perspective wrap but the network collapsed with this type of augmentation. Probably it was too \"aggressive\" and made the job of generator more difficult. I think that I'll carry out more experiments. Also, I've seen somewhere a commentary that rotations could undermine the performance as well.</p>\n\n<p>Maybe one just should apply the augmentations more carefully, or pick another architecture. I didn't experiment with it too much.</p>",
      "rawMarkdown": "I also would be glad to know more. The only thing I am using right now is horizontal flip and cropping. I tried perspective wrap but the network collapsed with this type of augmentation. Probably it was too \"aggressive\" and made the job of generator more difficult. I think that I'll carry out more experiments. Also, I've seen somewhere a commentary that rotations could undermine the performance as well.\n\nMaybe one just should apply the augmentations more carefully, or pick another architecture. I didn't experiment with it too much.",
      "votes": null
    },
    {
      "id": "591272",
      "postDate": "08/03/2019 12:43:14",
      "content": "<p>so weird because when I add rotation, it boost my score a bit :)</p>",
      "rawMarkdown": "so weird because when I add rotation, it boost my score a bit :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 591230,
      "author_name": "purplejester",
      "author_url": "",
      "post_date": "08/03/2019 11:01:59",
      "content": "<p>I also would be glad to know more. The only thing I am using right now is horizontal flip and cropping. I tried perspective wrap but the network collapsed with this type of augmentation. Probably it was too \"aggressive\" and made the job of generator more difficult. I think that I'll carry out more experiments. Also, I've seen somewhere a commentary that rotations could undermine the performance as well.</p>\n\n<p>Maybe one just should apply the augmentations more carefully, or pick another architecture. I didn't experiment with it too much.</p>",
      "votes": null,
      "replies": [
        {
          "id": 591272,
          "author_name": "yangsaewon",
          "author_url": "",
          "post_date": "08/03/2019 12:43:14",
          "content": "<p>so weird because when I add rotation, it boost my score a bit :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "591050": "I haven't tried many augmentations, but I want to know what others think about the augmenations.\n\nI've tried with fixed augmentation as resize and centercrop (64 size)\n\nand used horizontal flip + random rotation as baseline\n\nhow are you guys dealing with this problem?\n\nare the augmentations really helpful for training GANs?\n\nI would really appreciate it if you share your experiments, thanks!",
    "591230": "I also would be glad to know more. The only thing I am using right now is horizontal flip and cropping. I tried perspective wrap but the network collapsed with this type of augmentation. Probably it was too \"aggressive\" and made the job of generator more difficult. I think that I'll carry out more experiments. Also, I've seen somewhere a commentary that rotations could undermine the performance as well.\n\nMaybe one just should apply the augmentations more carefully, or pick another architecture. I didn't experiment with it too much.",
    "591272": "so weird because when I add rotation, it boost my score a bit :)"
  },
  "source": "meta"
}