{
  "id": 101441,
  "title": "Queries regarding mode collapse in GAN",
  "url": "/competitions/generative-dog-images/discussion/101441",
  "author_name": "",
  "post_date": "2019-07-25T20:14:03.932409400Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Did anyone came across the condition of Mode Collapse while training their GAN. How did you tackle it?</p>\n\n<p>I trained my DCGAN for 40 epochs and I think my GAN is suffering from this condition, because after the training when I generated some images to visualize them they all looked very similar with very minimal changes. I have used BatchNorm, Dropout layers, LeakyRELu as activation and ADAM as optimizer.</p>\n\n<p>Does anyone has any tips on this?</p>",
  "messages": [
    {
      "id": "584359",
      "postDate": "07/25/2019 20:14:03",
      "content": "<p>Did anyone came across the condition of Mode Collapse while training their GAN. How did you tackle it?</p>\n\n<p>I trained my DCGAN for 40 epochs and I think my GAN is suffering from this condition, because after the training when I generated some images to visualize them they all looked very similar with very minimal changes. I have used BatchNorm, Dropout layers, LeakyRELu as activation and ADAM as optimizer.</p>\n\n<p>Does anyone has any tips on this?</p>",
      "rawMarkdown": "Did anyone came across the condition of Mode Collapse while training their GAN. How did you tackle it?\n\nI trained my DCGAN for 40 epochs and I think my GAN is suffering from this condition, because after the training when I generated some images to visualize them they all looked very similar with very minimal changes. I have used BatchNorm, Dropout layers, LeakyRELu as activation and ADAM as optimizer.\n\nDoes anyone has any tips on this?",
      "votes": null
    },
    {
      "id": "584366",
      "postDate": "07/25/2019 20:31:53",
      "content": "<p>I had this a few times, when I used relativistic loss with TTUR. I had to set equal learning rates for D and G, and that fixed the problem.</p>",
      "rawMarkdown": "I had this a few times, when I used relativistic loss with TTUR. I had to set equal learning rates for D and G, and that fixed the problem.",
      "votes": null
    },
    {
      "id": "584387",
      "postDate": "07/25/2019 21:15:10",
      "content": "<p>I had this a few time with normal DCGAN and RaLSGAN . I had to change the learning rate . For e.g with my architecture when I set LR_G = 0.0003 and LR_D=0.00008 . Suddenly mode collapses . Then i tweak the LR_D to 0.0001 , and it worked .</p>",
      "rawMarkdown": "I had this a few time with normal DCGAN and RaLSGAN . I had to change the learning rate . For e.g with my architecture when I set LR_G = 0.0003 and LR_D=0.00008 . Suddenly mode collapses . Then i tweak the LR_D to 0.0001 , and it worked .",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 584366,
      "author_name": "dvorobiev",
      "author_url": "",
      "post_date": "07/25/2019 20:31:53",
      "content": "<p>I had this a few times, when I used relativistic loss with TTUR. I had to set equal learning rates for D and G, and that fixed the problem.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 584387,
      "author_name": "phoenix9032",
      "author_url": "",
      "post_date": "07/25/2019 21:15:10",
      "content": "<p>I had this a few time with normal DCGAN and RaLSGAN . I had to change the learning rate . For e.g with my architecture when I set LR_G = 0.0003 and LR_D=0.00008 . Suddenly mode collapses . Then i tweak the LR_D to 0.0001 , and it worked .</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "584359": "Did anyone came across the condition of Mode Collapse while training their GAN. How did you tackle it?\n\nI trained my DCGAN for 40 epochs and I think my GAN is suffering from this condition, because after the training when I generated some images to visualize them they all looked very similar with very minimal changes. I have used BatchNorm, Dropout layers, LeakyRELu as activation and ADAM as optimizer.\n\nDoes anyone has any tips on this?",
    "584366": "I had this a few times, when I used relativistic loss with TTUR. I had to set equal learning rates for D and G, and that fixed the problem.",
    "584387": "I had this a few time with normal DCGAN and RaLSGAN . I had to change the learning rate . For e.g with my architecture when I set LR_G = 0.0003 and LR_D=0.00008 . Suddenly mode collapses . Then i tweak the LR_D to 0.0001 , and it worked ."
  },
  "source": "meta"
}