{
  "id": 480775,
  "title": "TTA not helping?",
  "url": "/competitions/hms-harmful-brain-activity-classification/discussion/480775",
  "author_name": "Cody_Null",
  "post_date": "2024-02-29T23:35:49.297000",
  "votes": 8,
  "comment_count": 14,
  "views": 0,
  "content": "<p>I have now built several models on different architectures and have yet to find a solution where Test Time Augmentations (TTA) preform better on LB. Has anyone else seen this? Does TTA work for anyone? </p>",
  "messages": [
    {
      "id": 2675467,
      "postDate": "2024-02-29T23:35:49.297Z",
      "content": "<p>I have now built several models on different architectures and have yet to find a solution where Test Time Augmentations (TTA) preform better on LB. Has anyone else seen this? Does TTA work for anyone? </p>",
      "rawMarkdown": "I have now built several models on different architectures and have yet to find a solution where Test Time Augmentations (TTA) preform better on LB. Has anyone else seen this? Does TTA work for anyone? ",
      "votes": 8
    },
    {
      "id": 2675584,
      "postDate": "2024-03-01T02:37:43.313Z",
      "content": "<p>Data augmentation has only had a negative impact on my results (currently)</p>",
      "rawMarkdown": "Data augmentation has only had a negative impact on my results (currently)",
      "votes": 2,
      "replies": [
        {
          "id": 2675592,
          "postDate": "2024-03-01T02:51:32.407Z",
          "content": "<p>Though it took some time I am getting positive effect on CV from augmentations but not in test time</p>",
          "rawMarkdown": "Though it took some time I am getting positive effect on CV from augmentations but not in test time",
          "votes": 2
        },
        {
          "id": 2675703,
          "postDate": "2024-03-01T04:42:31.113Z",
          "content": "<p>Did you reach 0.26 without augmentations?</p>",
          "rawMarkdown": "Did you reach 0.26 without augmentations?",
          "replies": [
            {
              "id": 2675818,
              "postDate": "2024-03-01T05:49:55.917Z",
              "content": "<p>Yes. If using mixup, the result is 0.27.<br>\nOther data augmentations did not have a positive impact, whether it was CV or LB.</p>",
              "rawMarkdown": "Yes. If using mixup, the result is 0.27.\nOther data augmentations did not have a positive impact, whether it was CV or LB.",
              "votes": 5
            }
          ]
        },
        {
          "id": 2676269,
          "postDate": "2024-03-01T11:30:50.330Z",
          "content": "<p>If you don’t mind, is this the result obtained using RAW EGG?</p>",
          "rawMarkdown": "If you don’t mind, is this the result obtained using RAW EGG?",
          "votes": 1,
          "replies": [
            {
              "id": 2676327,
              "postDate": "2024-03-01T12:36:56.220Z",
              "content": "<p>The combination of time-frequency maps (spec.) and raw EEG as multiple inputs for a single model.</p>",
              "rawMarkdown": "The combination of time-frequency maps (spec.) and raw EEG as multiple inputs for a single model.",
              "votes": 16
            },
            {
              "id": 2676339,
              "postDate": "2024-03-01T12:48:55.613Z",
              "content": "<p>great works!</p>",
              "rawMarkdown": "great works!"
            },
            {
              "id": 2676386,
              "postDate": "2024-03-01T13:29:13.060Z",
              "content": "<p>Do you have model with two encoders, for spec. and raw EEG?</p>",
              "rawMarkdown": "Do you have model with two encoders, for spec. and raw EEG?"
            },
            {
              "id": 2677211,
              "postDate": "2024-03-02T03:18:22.630Z",
              "content": "<p>Yes, the backbone is EfficientNet-B0</p>",
              "rawMarkdown": "Yes, the backbone is EfficientNet-B0",
              "votes": 1
            }
          ]
        }
      ]
    },
    {
      "id": 2676315,
      "postDate": "2024-03-01T12:19:59.310Z",
      "content": "<p>TTA has effect -0.01 on LB.<br>\nI think it's import to use very simple augmentation in test time inference because complicated augmentation may make model more confusing than simple one.</p>",
      "rawMarkdown": "TTA has effect -0.01 on LB.\nI think it's import to use very simple augmentation in test time inference because complicated augmentation may make model more confusing than simple one.\n"
    },
    {
      "id": 2676276,
      "postDate": "2024-03-01T11:39:31.850Z",
      "content": "<p>wow, i've never thought TTA! Thanks for sharing your good idea😀</p>",
      "rawMarkdown": "wow, i've never thought TTA! Thanks for sharing your good idea😀"
    },
    {
      "id": 2676271,
      "postDate": "2024-03-01T11:32:14.717Z",
      "content": "<p>TTA has no effect on LB</p>",
      "rawMarkdown": "TTA has no effect on LB"
    },
    {
      "id": 2677192,
      "postDate": "2024-03-02T02:12:44.597Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 2677212,
          "postDate": "2024-03-02T03:19:19.317Z",
          "content": "<p>How about CV</p>",
          "rawMarkdown": "How about CV",
          "votes": 1
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2675584,
      "author_name": "Yuri Sun",
      "author_url": "",
      "post_date": "2024-03-01T02:37:43.313000",
      "content": "<p>Data augmentation has only had a negative impact on my results (currently)</p>",
      "votes": 2,
      "replies": [
        {
          "id": 2675592,
          "author_name": "Cody_Null",
          "author_url": "",
          "post_date": "2024-03-01T02:51:32.407000",
          "content": "<p>Though it took some time I am getting positive effect on CV from augmentations but not in test time</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 2675703,
          "author_name": "Gunes Evitan",
          "author_url": "",
          "post_date": "2024-03-01T04:42:31.113000",
          "content": "<p>Did you reach 0.26 without augmentations?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2675818,
              "author_name": "Yuri Sun",
              "author_url": "",
              "post_date": "2024-03-01T05:49:55.917000",
              "content": "<p>Yes. If using mixup, the result is 0.27.<br>\nOther data augmentations did not have a positive impact, whether it was CV or LB.</p>",
              "votes": 5,
              "replies": []
            }
          ]
        },
        {
          "id": 2676269,
          "author_name": "Donghui Zhang",
          "author_url": "",
          "post_date": "2024-03-01T11:30:50.330000",
          "content": "<p>If you don’t mind, is this the result obtained using RAW EGG?</p>",
          "votes": 1,
          "replies": [
            {
              "id": 2676327,
              "author_name": "Yuri Sun",
              "author_url": "",
              "post_date": "2024-03-01T12:36:56.220000",
              "content": "<p>The combination of time-frequency maps (spec.) and raw EEG as multiple inputs for a single model.</p>",
              "votes": 16,
              "replies": []
            },
            {
              "id": 2676339,
              "author_name": "Donghui Zhang",
              "author_url": "",
              "post_date": "2024-03-01T12:48:55.613000",
              "content": "<p>great works!</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2676386,
              "author_name": "Georgii Aparin",
              "author_url": "",
              "post_date": "2024-03-01T13:29:13.060000",
              "content": "<p>Do you have model with two encoders, for spec. and raw EEG?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2677211,
              "author_name": "Yuri Sun",
              "author_url": "",
              "post_date": "2024-03-02T03:18:22.630000",
              "content": "<p>Yes, the backbone is EfficientNet-B0</p>",
              "votes": 1,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2676315,
      "author_name": "Aurora_blue",
      "author_url": "",
      "post_date": "2024-03-01T12:19:59.310000",
      "content": "<p>TTA has effect -0.01 on LB.<br>\nI think it's import to use very simple augmentation in test time inference because complicated augmentation may make model more confusing than simple one.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2676276,
      "author_name": "Peter",
      "author_url": "",
      "post_date": "2024-03-01T11:39:31.850000",
      "content": "<p>wow, i've never thought TTA! Thanks for sharing your good idea😀</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2676271,
      "author_name": "Donghui Zhang",
      "author_url": "",
      "post_date": "2024-03-01T11:32:14.717000",
      "content": "<p>TTA has no effect on LB</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2677192,
      "author_name": "",
      "author_url": "",
      "post_date": "2024-03-02T02:12:44.597000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 2677212,
          "author_name": "Yuri Sun",
          "author_url": "",
          "post_date": "2024-03-02T03:19:19.317000",
          "content": "<p>How about CV</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2675467": "I have now built several models on different architectures and have yet to find a solution where Test Time Augmentations (TTA) preform better on LB. Has anyone else seen this? Does TTA work for anyone? ",
    "2675584": "Data augmentation has only had a negative impact on my results (currently)",
    "2676315": "TTA has effect -0.01 on LB.\nI think it's import to use very simple augmentation in test time inference because complicated augmentation may make model more confusing than simple one.\n",
    "2676276": "wow, i've never thought TTA! Thanks for sharing your good idea😀",
    "2676271": "TTA has no effect on LB",
    "2677192": ""
  }
}