{
  "id": 292849,
  "title": "Experimental GAN approach, model too cautious, ideas?",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/292849",
  "author_name": "oliver",
  "post_date": "2021-12-03T18:05:08.188000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi, I'm trying an experimental GAN approach here, not unlike an Autoencoder. <br>\nI've got some decent (nothing great) results, but I've noticed that the model is rather cautious, i.e. it only outputs the masks where it's very confident but rather outputs nothing when not sure.</p>\n<p>I've googled a great deal, but I'm not quite sure how I'd penalize its cautiousness?<br>\nUsing simple bce and l1.loss with the mask at the moment.<br>\nThanks for any ideas I could try.</p>\n<p>I also tried adding a perceptual loss, results are similar.</p>",
  "messages": [
    {
      "id": 1604814,
      "postDate": "2021-12-03T18:05:08.190Z",
      "content": "<p>Hi, I'm trying an experimental GAN approach here, not unlike an Autoencoder. <br>\nI've got some decent (nothing great) results, but I've noticed that the model is rather cautious, i.e. it only outputs the masks where it's very confident but rather outputs nothing when not sure.</p>\n<p>I've googled a great deal, but I'm not quite sure how I'd penalize its cautiousness?<br>\nUsing simple bce and l1.loss with the mask at the moment.<br>\nThanks for any ideas I could try.</p>\n<p>I also tried adding a perceptual loss, results are similar.</p>",
      "rawMarkdown": "Hi, I'm trying an experimental GAN approach here, not unlike an Autoencoder. \nI've got some decent (nothing great) results, but I've noticed that the model is rather cautious, i.e. it only outputs the masks where it's very confident but rather outputs nothing when not sure.\n\nI've googled a great deal, but I'm not quite sure how I'd penalize its cautiousness?\nUsing simple bce and l1.loss with the mask at the moment.\nThanks for any ideas I could try.\n\nI also tried adding a perceptual loss, results are similar.",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1604814": "Hi, I'm trying an experimental GAN approach here, not unlike an Autoencoder. \nI've got some decent (nothing great) results, but I've noticed that the model is rather cautious, i.e. it only outputs the masks where it's very confident but rather outputs nothing when not sure.\n\nI've googled a great deal, but I'm not quite sure how I'd penalize its cautiousness?\nUsing simple bce and l1.loss with the mask at the moment.\nThanks for any ideas I could try.\n\nI also tried adding a perceptual loss, results are similar."
  }
}