{
  "id": 668066,
  "title": "Does TTA hurt performance? (LB dropped 0.04)",
  "url": "/competitions/csiro-biomass/discussion/668066",
  "author_name": "hikaggler",
  "post_date": "2026-01-15T04:51:10.082000",
  "votes": 4,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n<p>I am currently working on an ensemble approach using DINOv2 and SigLIP based models. Recently, I tried applying Test Time Augmentation (TTA) to improve my inference robustness, but I encountered an unexpected result.</p>\n<p><strong>What I tried</strong>: I used the following TTA modes and averaged the predictions: tta_modes = ('original', 'hflip', 'vflip', 'both')</p>\n<p><strong>Result</strong>: My LB score dropped by approximately 0.04 compared to inference without TTA. Usually, TTA provides a slight boost or stays neutral, but this drop was significant.</p>\n<p><strong>Question</strong>: Has anyone else experienced a score drop when using TTA in this competition?</p>\n<p>Any insights would be appreciated!</p>\n<p>Thanks.</p>",
  "messages": [
    {
      "id": 3391506,
      "postDate": "2026-01-15T04:51:10.083Z",
      "content": "<p>Hi everyone,</p>\n<p>I am currently working on an ensemble approach using DINOv2 and SigLIP based models. Recently, I tried applying Test Time Augmentation (TTA) to improve my inference robustness, but I encountered an unexpected result.</p>\n<p><strong>What I tried</strong>: I used the following TTA modes and averaged the predictions: tta_modes = ('original', 'hflip', 'vflip', 'both')</p>\n<p><strong>Result</strong>: My LB score dropped by approximately 0.04 compared to inference without TTA. Usually, TTA provides a slight boost or stays neutral, but this drop was significant.</p>\n<p><strong>Question</strong>: Has anyone else experienced a score drop when using TTA in this competition?</p>\n<p>Any insights would be appreciated!</p>\n<p>Thanks.</p>",
      "rawMarkdown": "Hi everyone,\n\nI am currently working on an ensemble approach using DINOv2 and SigLIP based models. Recently, I tried applying Test Time Augmentation (TTA) to improve my inference robustness, but I encountered an unexpected result.\n\n**What I tried**: I used the following TTA modes and averaged the predictions: tta_modes = ('original', 'hflip', 'vflip', 'both')\n\n**Result**: My LB score dropped by approximately 0.04 compared to inference without TTA. Usually, TTA provides a slight boost or stays neutral, but this drop was significant.\n\n**Question**: Has anyone else experienced a score drop when using TTA in this competition?\n\nAny insights would be appreciated!\n\nThanks.",
      "votes": 4
    },
    {
      "id": 3391639,
      "postDate": "2026-01-15T11:22:28.877Z",
      "content": "<p>For me, flip TTA improves the score quite significantly. Did you train your models using the same flips as augmentation?</p>",
      "rawMarkdown": "For me, flip TTA improves the score quite significantly. Did you train your models using the same flips as augmentation?",
      "votes": 1,
      "replies": [
        {
          "id": 3392054,
          "postDate": "2026-01-16T08:34:56.087Z",
          "content": "<p>I used this transform for train data.</p>\n<p><code>def get_train_transforms():\n    return A.Compose([\n        A.Resize(config.img_size, config.img_size),\n        A.HorizontalFlip(p=0.5), A.VerticalFlip(p=0.5), A.RandomRotate90(p=0.5),\n        A.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.08, p=0.5),\n        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), ToTensorV2()\n    ])</code></p>",
          "rawMarkdown": "I used this transform for train data.\n\n`def get_train_transforms():\n    return A.Compose([\n        A.Resize(config.img_size, config.img_size),\n        A.HorizontalFlip(p=0.5), A.VerticalFlip(p=0.5), A.RandomRotate90(p=0.5),\n        A.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.08, p=0.5),\n        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), ToTensorV2()\n    ])`",
          "replies": [
            {
              "id": 3392132,
              "postDate": "2026-01-16T10:28:51.850Z",
              "content": "<p>Since it does have both horizontal and vertical flip in the augmentations, you should definitely be getting an increase in the score, not a decrease. \nI see that you are using a random rotate 90. Are you sure that you are doing this on a square, like left and right half of the image and not the entire image?\nAlso, make sure that when you are doing it in TTA, you are doing the flips on the entire image, or else their boundaries will not match.</p>",
              "rawMarkdown": "Since it does have both horizontal and vertical flip in the augmentations, you should definitely be getting an increase in the score, not a decrease. \nI see that you are using a random rotate 90. Are you sure that you are doing this on a square, like left and right half of the image and not the entire image?\nAlso, make sure that when you are doing it in TTA, you are doing the flips on the entire image, or else their boundaries will not match.",
              "votes": 2
            },
            {
              "id": 3392172,
              "postDate": "2026-01-16T11:43:46.950Z",
              "rawMarkdown": "",
              "votes": -3,
              "isDeleted": true
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 3391639,
      "author_name": "tennogh",
      "author_url": "",
      "post_date": "2026-01-15T11:22:28.877000",
      "content": "<p>For me, flip TTA improves the score quite significantly. Did you train your models using the same flips as augmentation?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3392054,
          "author_name": "hikaggler",
          "author_url": "",
          "post_date": "2026-01-16T08:34:56.087000",
          "content": "<p>I used this transform for train data.</p>\n<p><code>def get_train_transforms():\n    return A.Compose([\n        A.Resize(config.img_size, config.img_size),\n        A.HorizontalFlip(p=0.5), A.VerticalFlip(p=0.5), A.RandomRotate90(p=0.5),\n        A.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.08, p=0.5),\n        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), ToTensorV2()\n    ])</code></p>",
          "votes": 0,
          "replies": [
            {
              "id": 3392132,
              "author_name": "Komil Parmar",
              "author_url": "",
              "post_date": "2026-01-16T10:28:51.850000",
              "content": "<p>Since it does have both horizontal and vertical flip in the augmentations, you should definitely be getting an increase in the score, not a decrease. \nI see that you are using a random rotate 90. Are you sure that you are doing this on a square, like left and right half of the image and not the entire image?\nAlso, make sure that when you are doing it in TTA, you are doing the flips on the entire image, or else their boundaries will not match.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3392172,
              "author_name": "",
              "author_url": "",
              "post_date": "2026-01-16T11:43:46.950000",
              "content": "",
              "votes": -3,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3391506": "Hi everyone,\n\nI am currently working on an ensemble approach using DINOv2 and SigLIP based models. Recently, I tried applying Test Time Augmentation (TTA) to improve my inference robustness, but I encountered an unexpected result.\n\n**What I tried**: I used the following TTA modes and averaged the predictions: tta_modes = ('original', 'hflip', 'vflip', 'both')\n\n**Result**: My LB score dropped by approximately 0.04 compared to inference without TTA. Usually, TTA provides a slight boost or stays neutral, but this drop was significant.\n\n**Question**: Has anyone else experienced a score drop when using TTA in this competition?\n\nAny insights would be appreciated!\n\nThanks.",
    "3391639": "For me, flip TTA improves the score quite significantly. Did you train your models using the same flips as augmentation?"
  }
}