{
  "id": 399140,
  "title": "Can we use pretrained models?",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/399140",
  "author_name": "",
  "post_date": "2023-04-02T16:11:50.565304Z",
  "votes": 6,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Just want to confirm since I do not see this in rules, if using pretrained models is allowed?</p>",
  "messages": [
    {
      "id": "2206528",
      "postDate": "04/02/2023 16:11:50",
      "content": "<p>Just want to confirm since I do not see this in rules, if using pretrained models is allowed?</p>",
      "rawMarkdown": "Just want to confirm since I do not see this in rules, if using pretrained models is allowed?",
      "votes": null
    },
    {
      "id": "2206556",
      "postDate": "04/02/2023 16:40:31",
      "content": "<p>Yes, but we have to be able to reproduce your work if you win (e.g. you have to show us how you did the pretraining and we should be able to run it ourselves).</p>",
      "rawMarkdown": "Yes, but we have to be able to reproduce your work if you win (e.g. you have to show us how you did the pretraining and we should be able to run it ourselves).",
      "votes": null
    },
    {
      "id": "2206791",
      "postDate": "04/02/2023 21:00:08",
      "content": "<p>Thanks. I meant, can we use publicly available pretrained models which have already been trained on other datasets?</p>\n<p>For example, a UNet model with resnet backbone which is pretrained on imagenet dataset.<br>\n<a href=\"https://github.com/qubvel/segmentation_models#models-and-backbones\" target=\"_blank\">https://github.com/qubvel/segmentation_models#models-and-backbones</a></p>",
      "rawMarkdown": "Thanks. I meant, can we use publicly available pretrained models which have already been trained on other datasets?\n\nFor example, a UNet model with resnet backbone which is pretrained on imagenet dataset.\nhttps://github.com/qubvel/segmentation_models#models-and-backbones",
      "votes": null
    },
    {
      "id": "2206849",
      "postDate": "04/03/2023 00:09:30",
      "content": "<p>Sure! Again, that's all cool as long as we can reproduce it if you win. </p>",
      "rawMarkdown": "Sure! Again, that's all cool as long as we can reproduce it if you win.",
      "votes": null
    },
    {
      "id": "2207623",
      "postDate": "04/03/2023 14:24:53",
      "content": "<p>Thanks for confirming </p>",
      "rawMarkdown": "Thanks for confirming",
      "votes": null
    },
    {
      "id": "2251148",
      "postDate": "05/09/2023 05:52:24",
      "content": "<p><a href=\"https://www.kaggle.com/jpposma\" target=\"_blank\">@jpposma</a> Can I ask more precise about \"reproducibility\"?</p>\n<p>Lots of pre-trained models uses research-only datasets like ImageNet, MS CoCo etc.. So does \"reproducible\" means we should have access right to these datasets?</p>\n<p>If it is yes, this makes using pre-trained weights almost impossible. So the next question is why this regulation exists on this competition?</p>",
      "rawMarkdown": "jpposma Can I ask more precise about \"reproducibility\"?\n\nLots of pre-trained models uses research-only datasets like ImageNet, MS CoCo etc.. So does \"reproducible\" means we should have access right to these datasets?\n\nIf it is yes, this makes using pre-trained weights almost impossible. So the next question is why this regulation exists on this competition?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2206556,
      "author_name": "jpposma",
      "author_url": "",
      "post_date": "04/02/2023 16:40:31",
      "content": "<p>Yes, but we have to be able to reproduce your work if you win (e.g. you have to show us how you did the pretraining and we should be able to run it ourselves).</p>",
      "votes": null,
      "replies": [
        {
          "id": 2206791,
          "author_name": "vedkumarpatel",
          "author_url": "",
          "post_date": "04/02/2023 21:00:08",
          "content": "<p>Thanks. I meant, can we use publicly available pretrained models which have already been trained on other datasets?</p>\n<p>For example, a UNet model with resnet backbone which is pretrained on imagenet dataset.<br>\n<a href=\"https://github.com/qubvel/segmentation_models#models-and-backbones\" target=\"_blank\">https://github.com/qubvel/segmentation_models#models-and-backbones</a></p>",
          "votes": null,
          "replies": [
            {
              "id": 2206849,
              "author_name": "jpposma",
              "author_url": "",
              "post_date": "04/03/2023 00:09:30",
              "content": "<p>Sure! Again, that's all cool as long as we can reproduce it if you win. </p>",
              "votes": null,
              "replies": [
                {
                  "id": 2207623,
                  "author_name": "vedkumarpatel",
                  "author_url": "",
                  "post_date": "04/03/2023 14:24:53",
                  "content": "<p>Thanks for confirming </p>",
                  "votes": null,
                  "replies": []
                },
                {
                  "id": 2251148,
                  "author_name": "tatamikenn",
                  "author_url": "",
                  "post_date": "05/09/2023 05:52:24",
                  "content": "<p><a href=\"https://www.kaggle.com/jpposma\" target=\"_blank\">@jpposma</a> Can I ask more precise about \"reproducibility\"?</p>\n<p>Lots of pre-trained models uses research-only datasets like ImageNet, MS CoCo etc.. So does \"reproducible\" means we should have access right to these datasets?</p>\n<p>If it is yes, this makes using pre-trained weights almost impossible. So the next question is why this regulation exists on this competition?</p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2206528": "Just want to confirm since I do not see this in rules, if using pretrained models is allowed?",
    "2206556": "Yes, but we have to be able to reproduce your work if you win (e.g. you have to show us how you did the pretraining and we should be able to run it ourselves).",
    "2206791": "Thanks. I meant, can we use publicly available pretrained models which have already been trained on other datasets?\n\nFor example, a UNet model with resnet backbone which is pretrained on imagenet dataset.\nhttps://github.com/qubvel/segmentation_models#models-and-backbones",
    "2206849": "Sure! Again, that's all cool as long as we can reproduce it if you win.",
    "2207623": "Thanks for confirming",
    "2251148": "jpposma Can I ask more precise about \"reproducibility\"?\n\nLots of pre-trained models uses research-only datasets like ImageNet, MS CoCo etc.. So does \"reproducible\" means we should have access right to these datasets?\n\nIf it is yes, this makes using pre-trained weights almost impossible. So the next question is why this regulation exists on this competition?"
  },
  "source": "meta"
}