{
  "id": 472167,
  "title": "Transformer architecture",
  "url": "/competitions/hms-harmful-brain-activity-classification/discussion/472167",
  "author_name": "MiHu",
  "post_date": "2024-01-31T03:08:53.728000",
  "votes": 7,
  "comment_count": 9,
  "views": 0,
  "content": "<p>Most of the architectures I have seen so far are based on CNN (EfficientNet). Has anyone tried transformer architecture or anything else</p>",
  "messages": [
    {
      "id": 2627995,
      "postDate": "2024-01-31T03:08:53.730Z",
      "content": "<p>Most of the architectures I have seen so far are based on CNN (EfficientNet). Has anyone tried transformer architecture or anything else</p>",
      "rawMarkdown": "Most of the architectures I have seen so far are based on CNN (EfficientNet). Has anyone tried transformer architecture or anything else\n",
      "votes": 7
    },
    {
      "id": 2629606,
      "postDate": "2024-01-31T22:32:39.413Z",
      "content": "<p>I think this may just be because it is a leader in the public models. I am certain many other options would preform well. It also seems the Resnet34d is useful</p>",
      "rawMarkdown": "I think this may just be because it is a leader in the public models. I am certain many other options would preform well. It also seems the Resnet34d is useful",
      "votes": 2,
      "replies": [
        {
          "id": 2629610,
          "postDate": "2024-01-31T22:38:50.923Z",
          "content": "<p>As far as I've been trying, I can say that Mobilenetv3 did not perform very well. Will have a look tomorrow on how densenet 169 performed.</p>",
          "rawMarkdown": "As far as I've been trying, I can say that Mobilenetv3 did not perform very well. Will have a look tomorrow on how densenet 169 performed.",
          "votes": 2,
          "replies": [
            {
              "id": 2629626,
              "postDate": "2024-01-31T22:53:56.127Z",
              "rawMarkdown": "",
              "isDeleted": true
            },
            {
              "id": 2630538,
              "postDate": "2024-02-01T10:42:33.257Z",
              "content": "<p>I tried MAXVIT and it didn't work as well as EfficientNet</p>",
              "rawMarkdown": "I tried MAXVIT and it didn't work as well as EfficientNet",
              "votes": 2
            },
            {
              "id": 2633299,
              "postDate": "2024-02-03T01:29:17.817Z",
              "content": "<p>Given the particularity of the waveform plot, it is felt that patch-based methods are not sufficient to capture this particularity</p>",
              "rawMarkdown": "Given the particularity of the waveform plot, it is felt that patch-based methods are not sufficient to capture this particularity",
              "votes": 1
            }
          ]
        }
      ]
    },
    {
      "id": 2629592,
      "postDate": "2024-01-31T22:21:20.123Z",
      "content": "<p>I wanted to try convnexttiny, but I'm very confident that it will not load in the memory.</p>",
      "rawMarkdown": "I wanted to try convnexttiny, but I'm very confident that it will not load in the memory.",
      "votes": 2,
      "replies": [
        {
          "id": 2632654,
          "postDate": "2024-02-02T14:25:32.397Z",
          "content": "<p>I tried convnextSmall, and it ran out of memory. I'm not sure if I tried the tiny one because it was a few weeks ago, but from what I remember, I tried it, it didn't run out of memory, and the result wasn't good.</p>",
          "rawMarkdown": "I tried convnextSmall, and it ran out of memory. I'm not sure if I tried the tiny one because it was a few weeks ago, but from what I remember, I tried it, it didn't run out of memory, and the result wasn't good.",
          "votes": 1,
          "replies": [
            {
              "id": 2632803,
              "postDate": "2024-02-02T16:31:22.467Z",
              "content": "<p>For me convnexttiny as well went OOM. Did you train on GPU?</p>",
              "rawMarkdown": "For me convnexttiny as well went OOM. Did you train on GPU?",
              "votes": 2
            },
            {
              "id": 2632851,
              "postDate": "2024-02-02T16:58:35.467Z",
              "content": "<p>Yes a use GPU in colab…</p>",
              "rawMarkdown": "Yes a use GPU in colab...",
              "votes": 2
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2629606,
      "author_name": "Cody_Null",
      "author_url": "",
      "post_date": "2024-01-31T22:32:39.413000",
      "content": "<p>I think this may just be because it is a leader in the public models. I am certain many other options would preform well. It also seems the Resnet34d is useful</p>",
      "votes": 2,
      "replies": [
        {
          "id": 2629610,
          "author_name": "stefanoclss",
          "author_url": "",
          "post_date": "2024-01-31T22:38:50.923000",
          "content": "<p>As far as I've been trying, I can say that Mobilenetv3 did not perform very well. Will have a look tomorrow on how densenet 169 performed.</p>",
          "votes": 2,
          "replies": [
            {
              "id": 2629626,
              "author_name": "",
              "author_url": "",
              "post_date": "2024-01-31T22:53:56.127000",
              "content": "",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2630538,
              "author_name": "MiHu",
              "author_url": "",
              "post_date": "2024-02-01T10:42:33.257000",
              "content": "<p>I tried MAXVIT and it didn't work as well as EfficientNet</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2633299,
              "author_name": "MiHu",
              "author_url": "",
              "post_date": "2024-02-03T01:29:17.817000",
              "content": "<p>Given the particularity of the waveform plot, it is felt that patch-based methods are not sufficient to capture this particularity</p>",
              "votes": 1,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2629592,
      "author_name": "stefanoclss",
      "author_url": "",
      "post_date": "2024-01-31T22:21:20.123000",
      "content": "<p>I wanted to try convnexttiny, but I'm very confident that it will not load in the memory.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 2632654,
          "author_name": "Rafael Zimmermann",
          "author_url": "",
          "post_date": "2024-02-02T14:25:32.397000",
          "content": "<p>I tried convnextSmall, and it ran out of memory. I'm not sure if I tried the tiny one because it was a few weeks ago, but from what I remember, I tried it, it didn't run out of memory, and the result wasn't good.</p>",
          "votes": 1,
          "replies": [
            {
              "id": 2632803,
              "author_name": "stefanoclss",
              "author_url": "",
              "post_date": "2024-02-02T16:31:22.467000",
              "content": "<p>For me convnexttiny as well went OOM. Did you train on GPU?</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2632851,
              "author_name": "Rafael Zimmermann",
              "author_url": "",
              "post_date": "2024-02-02T16:58:35.467000",
              "content": "<p>Yes a use GPU in colab…</p>",
              "votes": 2,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2627995": "Most of the architectures I have seen so far are based on CNN (EfficientNet). Has anyone tried transformer architecture or anything else\n",
    "2629606": "I think this may just be because it is a leader in the public models. I am certain many other options would preform well. It also seems the Resnet34d is useful",
    "2629592": "I wanted to try convnexttiny, but I'm very confident that it will not load in the memory."
  }
}