{
  "id": 99223,
  "title": "hey guys I need help with my DCGAN pytorch",
  "url": "/competitions/generative-dog-images/discussion/99223",
  "author_name": "",
  "post_date": "2019-07-09T16:11:25.310850900Z",
  "votes": null,
  "comment_count": 7,
  "views": 0,
  "content": "",
  "messages": [
    {
      "id": "571468",
      "postDate": "07/09/2019 16:11:25",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "571475",
      "postDate": "07/09/2019 16:19:39",
      "content": "<p>hi guys,am having problem with my generator  converging ...I keep on getting a constant generator loss for a long time ..there is something am not doing right any piece on advice on ..what can I do to improve my model..I'll highly appreciate.....Thanks in advance..   <a href=\"https://www.kaggle.com/risperbevalyn/dcgan\">https://www.kaggle.com/risperbevalyn/dcgan</a></p>",
      "rawMarkdown": "hi guys,am having problem with my generator  converging ...I keep on getting a constant generator loss for a long time ..there is something am not doing right any piece on advice on ..what can I do to improve my model..I'll highly appreciate.....Thanks in advance..   https://www.kaggle.com/risperbevalyn/dcgan",
      "votes": null
    },
    {
      "id": "571492",
      "postDate": "07/09/2019 17:02:30",
      "content": "<p>Hi <a href=\"/risperbevalyn\">@risperbevalyn</a>, constant generator loss is not problem itself, since generator and discriminator constantly competing against each other, they can get better and better without change of their losses, if like you would learn to play tennis with your friend for example and always lose half of matches (constant loss), but getting better and better, as your friend. What you should check is generated images are getting better. Inspect with your own eyes. </p>",
      "rawMarkdown": "Hi @risperbevalyn, constant generator loss is not problem itself, since generator and discriminator constantly competing against each other, they can get better and better without change of their losses, if like you would learn to play tennis with your friend for example and always lose half of matches (constant loss), but getting better and better, as your friend. What you should check is generated images are getting better. Inspect with your own eyes.",
      "votes": null
    },
    {
      "id": "571566",
      "postDate": "07/09/2019 18:37:11",
      "content": "<p>I have read somewhere that BCEWithLogitsLoss() and Sigmoid() doesn't work together. So you can use BCEWithLogitsLoss() without Sigmoid() or you can use Sigmoid() and BCELoss() . Is that the case for you ?</p>",
      "rawMarkdown": "I have read somewhere that BCEWithLogitsLoss() and Sigmoid() doesn't work together. So you can use BCEWithLogitsLoss() without Sigmoid() or you can use Sigmoid() and BCELoss() . Is that the case for you ?",
      "votes": null
    },
    {
      "id": "571617",
      "postDate": "07/09/2019 20:39:15",
      "content": "<p>BCEWithLogitsLoss() already has a sigmoid unit. It's basically BCELoss() + Sigmoid()\n<a href=\"https://discuss.pytorch.org/t/bceloss-vs-bcewithlogitsloss/33586\">source</a></p>",
      "rawMarkdown": "BCEWithLogitsLoss() already has a sigmoid unit. It's basically BCELoss() + Sigmoid()\n[source](https://discuss.pytorch.org/t/bceloss-vs-bcewithlogitsloss/33586)",
      "votes": null
    },
    {
      "id": "572794",
      "postDate": "07/11/2019 12:10:54",
      "content": "<p>thanks  a lot <a href=\"/phoenix9032\">@phoenix9032</a> and <a href=\"/timetraveller98\">@timetraveller98</a> ....my model improved</p>",
      "rawMarkdown": "thanks  a lot @phoenix9032 and @timetraveller98 ....my model improved",
      "votes": null
    },
    {
      "id": "572801",
      "postDate": "07/11/2019 12:19:24",
      "content": "<p>.....I have been retraining whenever I get constant  loss for a long time😅 😄 🙉   .. ooh I got it now... thanks a lot  <a href=\"/kostyaatarik\">@kostyaatarik</a> 😃 👍 ...</p>",
      "rawMarkdown": ".....I have been retraining whenever I get constant  loss for a long time😅 😄 🙉   .. ooh I got it now... thanks a lot  @kostyaatarik 😃 👍 ...",
      "votes": null
    },
    {
      "id": "572806",
      "postDate": "07/11/2019 12:25:22",
      "content": "<p>Constant losses are even good sign, if you've been training your GAN for a while and D- and G-losses are still have a lot of variance then something is going wrong, variance in losses from batch to batch is ok, but on average losses should become pretty stable after a few epochs.</p>",
      "rawMarkdown": "Constant losses are even good sign, if you've been training your GAN for a while and D- and G-losses are still have a lot of variance then something is going wrong, variance in losses from batch to batch is ok, but on average losses should become pretty stable after a few epochs.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 571475,
      "author_name": "risperbevalyn",
      "author_url": "",
      "post_date": "07/09/2019 16:19:39",
      "content": "<p>hi guys,am having problem with my generator  converging ...I keep on getting a constant generator loss for a long time ..there is something am not doing right any piece on advice on ..what can I do to improve my model..I'll highly appreciate.....Thanks in advance..   <a href=\"https://www.kaggle.com/risperbevalyn/dcgan\">https://www.kaggle.com/risperbevalyn/dcgan</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 571492,
      "author_name": "kostyaatarik",
      "author_url": "",
      "post_date": "07/09/2019 17:02:30",
      "content": "<p>Hi <a href=\"/risperbevalyn\">@risperbevalyn</a>, constant generator loss is not problem itself, since generator and discriminator constantly competing against each other, they can get better and better without change of their losses, if like you would learn to play tennis with your friend for example and always lose half of matches (constant loss), but getting better and better, as your friend. What you should check is generated images are getting better. Inspect with your own eyes. </p>",
      "votes": null,
      "replies": [
        {
          "id": 572801,
          "author_name": "risperbevalyn",
          "author_url": "",
          "post_date": "07/11/2019 12:19:24",
          "content": "<p>.....I have been retraining whenever I get constant  loss for a long time😅 😄 🙉   .. ooh I got it now... thanks a lot  <a href=\"/kostyaatarik\">@kostyaatarik</a> 😃 👍 ...</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 572806,
          "author_name": "kostyaatarik",
          "author_url": "",
          "post_date": "07/11/2019 12:25:22",
          "content": "<p>Constant losses are even good sign, if you've been training your GAN for a while and D- and G-losses are still have a lot of variance then something is going wrong, variance in losses from batch to batch is ok, but on average losses should become pretty stable after a few epochs.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 571566,
      "author_name": "phoenix9032",
      "author_url": "",
      "post_date": "07/09/2019 18:37:11",
      "content": "<p>I have read somewhere that BCEWithLogitsLoss() and Sigmoid() doesn't work together. So you can use BCEWithLogitsLoss() without Sigmoid() or you can use Sigmoid() and BCELoss() . Is that the case for you ?</p>",
      "votes": null,
      "replies": [
        {
          "id": 571617,
          "author_name": "timetraveller98",
          "author_url": "",
          "post_date": "07/09/2019 20:39:15",
          "content": "<p>BCEWithLogitsLoss() already has a sigmoid unit. It's basically BCELoss() + Sigmoid()\n<a href=\"https://discuss.pytorch.org/t/bceloss-vs-bcewithlogitsloss/33586\">source</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 572794,
          "author_name": "risperbevalyn",
          "author_url": "",
          "post_date": "07/11/2019 12:10:54",
          "content": "<p>thanks  a lot <a href=\"/phoenix9032\">@phoenix9032</a> and <a href=\"/timetraveller98\">@timetraveller98</a> ....my model improved</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "571468": "",
    "571475": "hi guys,am having problem with my generator  converging ...I keep on getting a constant generator loss for a long time ..there is something am not doing right any piece on advice on ..what can I do to improve my model..I'll highly appreciate.....Thanks in advance..   https://www.kaggle.com/risperbevalyn/dcgan",
    "571492": "Hi @risperbevalyn, constant generator loss is not problem itself, since generator and discriminator constantly competing against each other, they can get better and better without change of their losses, if like you would learn to play tennis with your friend for example and always lose half of matches (constant loss), but getting better and better, as your friend. What you should check is generated images are getting better. Inspect with your own eyes.",
    "571566": "I have read somewhere that BCEWithLogitsLoss() and Sigmoid() doesn't work together. So you can use BCEWithLogitsLoss() without Sigmoid() or you can use Sigmoid() and BCELoss() . Is that the case for you ?",
    "571617": "BCEWithLogitsLoss() already has a sigmoid unit. It's basically BCELoss() + Sigmoid()\n[source](https://discuss.pytorch.org/t/bceloss-vs-bcewithlogitsloss/33586)",
    "572794": "thanks  a lot @phoenix9032 and @timetraveller98 ....my model improved",
    "572801": ".....I have been retraining whenever I get constant  loss for a long time😅 😄 🙉   .. ooh I got it now... thanks a lot  @kostyaatarik 😃 👍 ...",
    "572806": "Constant losses are even good sign, if you've been training your GAN for a while and D- and G-losses are still have a lot of variance then something is going wrong, variance in losses from batch to batch is ok, but on average losses should become pretty stable after a few epochs."
  },
  "source": "meta"
}