{
  "id": 578415,
  "title": "Share with dfl loss for yolo",
  "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/578415",
  "author_name": "Rafał Pawłowski",
  "post_date": "2025-05-10T17:14:51.669000",
  "votes": 3,
  "comment_count": 11,
  "views": 0,
  "content": "<p>What is your CV dfl loss for yolo model? I am a bit late to this competition, but I think this loss can be important.</p>",
  "messages": [
    {
      "id": 3199884,
      "postDate": "2025-05-11T16:11:26.820Z",
      "content": "<p>I will share my experiment. The best LB dfl loss is below.<br>\n(Comparison using the same inference pipeline.)</p>\n<table>\n<thead>\n<tr>\n<th>LB</th>\n<th>DFL Loss(Val)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>.826</td>\n<td>1.9499</td>\n</tr>\n<tr>\n<td>.798</td>\n<td>1.9327</td>\n</tr>\n<tr>\n<td>.792</td>\n<td>1.9538</td>\n</tr>\n</tbody>\n</table>\n<p>As discussed elsewhere, picking the best LB is hard, especially on what metric. I'm beginning to think it's not a single metric.</p>",
      "rawMarkdown": "I will share my experiment. The best LB dfl loss is below.\n(Comparison using the same inference pipeline.)\n| LB | DFL Loss(Val) |\n| --- | --- |\n| .826 | 1.9499 |\n| .798 | 1.9327|\n| .792 | 1.9538|\n\nAs discussed elsewhere, picking the best LB is hard, especially on what metric. I'm beginning to think it's not a single metric.\n",
      "votes": 5,
      "replies": [
        {
          "id": 3199887,
          "postDate": "2025-05-11T16:16:00.567Z",
          "content": "<p>Interesting, you got so high loss with fantastic LB. I have 0.83 val loss with 0.672 LB. </p>",
          "rawMarkdown": "Interesting, you got so high loss with fantastic LB. I have 0.83 val loss with 0.672 LB. ",
          "votes": 2
        },
        {
          "id": 3199898,
          "postDate": "2025-05-11T16:40:06.700Z",
          "content": "<p>I hope you will publish these models after the competition is over. </p>",
          "rawMarkdown": "I hope you will publish these models after the competition is over. ",
          "votes": 1,
          "replies": [
            {
              "id": 3200499,
              "postDate": "2025-05-12T16:54:21.327Z",
              "content": "<p><a href=\"https://www.kaggle.com/fautei\" target=\"_blank\">@fautei</a>, Can I join to your team? I am training 3d detector now and building a filtering stage. </p>",
              "rawMarkdown": "@fautei, Can I join to your team? I am training 3d detector now and building a filtering stage. "
            },
            {
              "id": 3200544,
              "postDate": "2025-05-12T18:12:02.557Z",
              "content": "<p>You can't do this because there is a limit on the total number of submissiona from participants, and I spent 97% of them alone((</p>",
              "rawMarkdown": "You can't do this because there is a limit on the total number of submissiona from participants, and I spent 97% of them alone(("
            },
            {
              "id": 3200574,
              "postDate": "2025-05-12T18:41:35.607Z",
              "content": "<p>The rules state 'Team mergers are allowed and can be performed by the Team leader. In order to merge, the combined Team must have a total Submission count less than or equal to the maximum allowed as of the Team Merger Deadline. The maximum allowed is the number of Submissions per day multiplied by the number of days the competition has been running.'<br>\nI have 14 subs, you have 325 in total 339. 67 days * 5 = 335, so after 1 day without submitting we can merge. <br>\nIt implicates we can merge on Wednesday. </p>",
              "rawMarkdown": "The rules state 'Team mergers are allowed and can be performed by the Team leader. In order to merge, the combined Team must have a total Submission count less than or equal to the maximum allowed as of the Team Merger Deadline. The maximum allowed is the number of Submissions per day multiplied by the number of days the competition has been running.'\nI have 14 subs, you have 325 in total 339. 67 days * 5 = 335, so after 1 day without submitting we can merge. \nIt implicates we can merge on Wednesday. "
            },
            {
              "id": 3200645,
              "postDate": "2025-05-12T20:25:03.683Z",
              "content": "<p>*2 days because I have submissions that scoring now. But it might make sense, I think different point of view might make my solution better.</p>",
              "rawMarkdown": "*2 days because I have submissions that scoring now. But it might make sense, I think different point of view might make my solution better.",
              "votes": 2
            },
            {
              "id": 3204763,
              "postDate": "2025-05-18T18:46:36.660Z",
              "content": "<p>Give me hope , I have tried enough of yolo's and now I want to move to 2.5d / 2d+3d approach for fighting the domain shift (diff orientations in 3d not capturable with 2d models)  <br>\nEquivariant models (fast converging and possibly robust to scales) so are you guys LB with 3d model or similar?? </p>",
              "rawMarkdown": "Give me hope , I have tried enough of yolo's and now I want to move to 2.5d / 2d+3d approach for fighting the domain shift (diff orientations in 3d not capturable with 2d models)  \nEquivariant models (fast converging and possibly robust to scales) so are you guys LB with 3d model or similar?? ",
              "isDeleted": true
            }
          ]
        },
        {
          "id": 3201613,
          "postDate": "2025-05-14T06:28:51.023Z",
          "content": "<p>Could you please tell me what your iou settings are when training?</p>",
          "rawMarkdown": "Could you please tell me what your iou settings are when training?"
        },
        {
          "id": 3201840,
          "postDate": "2025-05-14T13:14:31.863Z",
          "content": "<p>Did you use additional data? I can't make use of the external data on yolo. Btw, I have 0.83 val loss with 0.806 LB.</p>",
          "rawMarkdown": "Did you use additional data? I can't make use of the external data on yolo. Btw, I have 0.83 val loss with 0.806 LB."
        }
      ]
    },
    {
      "id": 3199246,
      "postDate": "2025-05-10T17:14:51.670Z",
      "content": "<p>What is your CV dfl loss for yolo model? I am a bit late to this competition, but I think this loss can be important.</p>",
      "rawMarkdown": "What is your CV dfl loss for yolo model? I am a bit late to this competition, but I think this loss can be important.",
      "votes": 3
    },
    {
      "id": 3204019,
      "postDate": "2025-05-17T17:28:24.080Z",
      "content": "<table>\n<thead>\n<tr>\n<th>LB</th>\n<th>DFL</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>.825</td>\n<td>.93377</td>\n</tr>\n<tr>\n<td>.748</td>\n<td>.97973</td>\n</tr>\n<tr>\n<td>.684</td>\n<td>.96167</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "| LB | DFL |\n| --- | --- |\n| .825 | .93377 |\n| .748| .97973|\n| .684| .96167|",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 3199884,
      "author_name": "yukiZ",
      "author_url": "",
      "post_date": "2025-05-11T16:11:26.820000",
      "content": "<p>I will share my experiment. The best LB dfl loss is below.<br>\n(Comparison using the same inference pipeline.)</p>\n<table>\n<thead>\n<tr>\n<th>LB</th>\n<th>DFL Loss(Val)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>.826</td>\n<td>1.9499</td>\n</tr>\n<tr>\n<td>.798</td>\n<td>1.9327</td>\n</tr>\n<tr>\n<td>.792</td>\n<td>1.9538</td>\n</tr>\n</tbody>\n</table>\n<p>As discussed elsewhere, picking the best LB is hard, especially on what metric. I'm beginning to think it's not a single metric.</p>",
      "votes": 5,
      "replies": [
        {
          "id": 3199887,
          "author_name": "Rafał Pawłowski",
          "author_url": "",
          "post_date": "2025-05-11T16:16:00.567000",
          "content": "<p>Interesting, you got so high loss with fantastic LB. I have 0.83 val loss with 0.672 LB. </p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 3199898,
          "author_name": "Maxim Ilyin",
          "author_url": "",
          "post_date": "2025-05-11T16:40:06.700000",
          "content": "<p>I hope you will publish these models after the competition is over. </p>",
          "votes": 1,
          "replies": [
            {
              "id": 3200499,
              "author_name": "Rafał Pawłowski",
              "author_url": "",
              "post_date": "2025-05-12T16:54:21.327000",
              "content": "<p><a href=\"https://www.kaggle.com/fautei\" target=\"_blank\">@fautei</a>, Can I join to your team? I am training 3d detector now and building a filtering stage. </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3200544,
              "author_name": "Maxim Ilyin",
              "author_url": "",
              "post_date": "2025-05-12T18:12:02.557000",
              "content": "<p>You can't do this because there is a limit on the total number of submissiona from participants, and I spent 97% of them alone((</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3200574,
              "author_name": "Rafał Pawłowski",
              "author_url": "",
              "post_date": "2025-05-12T18:41:35.607000",
              "content": "<p>The rules state 'Team mergers are allowed and can be performed by the Team leader. In order to merge, the combined Team must have a total Submission count less than or equal to the maximum allowed as of the Team Merger Deadline. The maximum allowed is the number of Submissions per day multiplied by the number of days the competition has been running.'<br>\nI have 14 subs, you have 325 in total 339. 67 days * 5 = 335, so after 1 day without submitting we can merge. <br>\nIt implicates we can merge on Wednesday. </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3200645,
              "author_name": "Maxim Ilyin",
              "author_url": "",
              "post_date": "2025-05-12T20:25:03.683000",
              "content": "<p>*2 days because I have submissions that scoring now. But it might make sense, I think different point of view might make my solution better.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3204763,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-05-18T18:46:36.660000",
              "content": "<p>Give me hope , I have tried enough of yolo's and now I want to move to 2.5d / 2d+3d approach for fighting the domain shift (diff orientations in 3d not capturable with 2d models)  <br>\nEquivariant models (fast converging and possibly robust to scales) so are you guys LB with 3d model or similar?? </p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3201613,
          "author_name": "Heeler-Deer",
          "author_url": "",
          "post_date": "2025-05-14T06:28:51.023000",
          "content": "<p>Could you please tell me what your iou settings are when training?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3201840,
          "author_name": "shanzhong8",
          "author_url": "",
          "post_date": "2025-05-14T13:14:31.863000",
          "content": "<p>Did you use additional data? I can't make use of the external data on yolo. Btw, I have 0.83 val loss with 0.806 LB.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3204019,
      "author_name": "MathieuD",
      "author_url": "",
      "post_date": "2025-05-17T17:28:24.080000",
      "content": "<table>\n<thead>\n<tr>\n<th>LB</th>\n<th>DFL</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>.825</td>\n<td>.93377</td>\n</tr>\n<tr>\n<td>.748</td>\n<td>.97973</td>\n</tr>\n<tr>\n<td>.684</td>\n<td>.96167</td>\n</tr>\n</tbody>\n</table>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3199884": "I will share my experiment. The best LB dfl loss is below.\n(Comparison using the same inference pipeline.)\n| LB | DFL Loss(Val) |\n| --- | --- |\n| .826 | 1.9499 |\n| .798 | 1.9327|\n| .792 | 1.9538|\n\nAs discussed elsewhere, picking the best LB is hard, especially on what metric. I'm beginning to think it's not a single metric.\n",
    "3199246": "What is your CV dfl loss for yolo model? I am a bit late to this competition, but I think this loss can be important.",
    "3204019": "| LB | DFL |\n| --- | --- |\n| .825 | .93377 |\n| .748| .97973|\n| .684| .96167|"
  }
}