{
  "owner": "labmlai",
  "repo": "annotated_deep_learning_paper_implementations",
  "number": 146,
  "html_url": "https://github.com/labmlai/annotated_deep_learning_paper_implementations/issues/146",
  "is_pull_request": false,
  "state": "closed",
  "state_reason": "completed",
  "title": "Bug in SA for DDPM UNet?",
  "author": "FutureXiang",
  "created_at": "2022-09-07T17:28:50Z",
  "updated_at": "2022-10-13T11:16:47Z",
  "closed_at": "2022-10-13T11:16:47Z",
  "labels": [
    "bug"
  ],
  "milestone": null,
  "comments_count": 3,
  "reactions": {
    "total_count": 0,
    "+1": 0,
    "-1": 0,
    "laugh": 0,
    "hooray": 0,
    "confused": 0,
    "heart": 0,
    "rocket": 0,
    "eyes": 0
  },
  "resolution_days": 35.74,
  "fix": {
    "closing_commit": null,
    "linked_prs": [],
    "best_guess_fix_commit": null,
    "has_fix": false
  },
  "fetched_at": "2026-07-28T12:22:15.889126+00:00",
  "comments": [
    {
      "author": "vpj",
      "created_at": "2022-09-11T12:02:35Z",
      "body": "Thanks, you are right! That's a typo and a big bug!"
    },
    {
      "author": "FutureXiang",
      "created_at": "2022-09-11T14:04:07Z",
      "body": "Thank you for your response!\r\n\r\nI compare the FID results on classifier-free guidance conditional CIFAR-10 with `attn = attn.softmax(dim=1)` and `attn = attn.softmax(dim=2)`, following the settings in the original DDPM paper.\r\n\r\nHowever, I observe **no difference** between the bugged and the correct model (FID \u00b10.05),\r\nand the correct model performs **even worse** than the bugged one when trained with more iterations (FID -0.02~0.3).\r\n\r\nHave any idea why it happens?"
    },
    {
      "author": "vpj",
      "created_at": "2022-09-12T03:11:06Z",
      "body": "This is strange. I guess the wrong softmax also provides a similar non-linearity to the correct softmax and gradient descent finds a way to use it. But I don't understand how the wrong softmax becomes better than the correct one. Wonder how well no attention will perform. I will also try to run some tests.\r\n\r\nBtw, I pushed the fix https://github.com/labmlai/annotated_deep_learning_paper_implementations/commit/7d1550dd67ef903959a6367165e17c15b89da5c8"
    }
  ]
}