{
  "id": 112570,
  "title": "Augmentations Strategies for this Competition. TTA?",
  "url": "/competitions/understanding_cloud_organization/discussion/112570",
  "author_name": "",
  "post_date": "2019-10-13T23:25:00.201407200Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I see most of the public kernels have used flip and rotate augmentations. Although they work well with most of the image datasets, I feel we have a unique situation here and does require special consideration for this dataset.</p>\n\n<p>Given we are dealing with <strong>Satellite images lighting (Solarize Augmentation)</strong> might play a key role in balancing the dataset? <em>I did try a couple of combinations and resulted in 0.659, so it seems worth trying.</em></p>\n\n<p>What other augmentation you feel could be important in this dataset? Although I haven't tried what's difference b/w TTA vs non-TTA models? Looking for insightful discussion. TIA</p>",
  "messages": [
    {
      "id": "648245",
      "postDate": "10/13/2019 23:25:00",
      "content": "<p>I see most of the public kernels have used flip and rotate augmentations. Although they work well with most of the image datasets, I feel we have a unique situation here and does require special consideration for this dataset.</p>\n\n<p>Given we are dealing with <strong>Satellite images lighting (Solarize Augmentation)</strong> might play a key role in balancing the dataset? <em>I did try a couple of combinations and resulted in 0.659, so it seems worth trying.</em></p>\n\n<p>What other augmentation you feel could be important in this dataset? Although I haven't tried what's difference b/w TTA vs non-TTA models? Looking for insightful discussion. TIA</p>",
      "rawMarkdown": "I see most of the public kernels have used flip and rotate augmentations. Although they work well with most of the image datasets, I feel we have a unique situation here and does require special consideration for this dataset.\n\nGiven we are dealing with **Satellite images lighting (Solarize Augmentation)** might play a key role in balancing the dataset? *I did try a couple of combinations and resulted in 0.659, so it seems worth trying.*\n\nWhat other augmentation you feel could be important in this dataset? Although I haven't tried what's difference b/w TTA vs non-TTA models? Looking for insightful discussion. TIA",
      "votes": null
    },
    {
      "id": "648849",
      "postDate": "10/14/2019 17:28:21",
      "content": "<p><a href=\"/vivekwisdom\">@vivekwisdom</a>  - What threshold did you use or ranges of threshold?</p>",
      "rawMarkdown": "vivekwisdom  - What threshold did you use or ranges of threshold?",
      "votes": null
    },
    {
      "id": "648904",
      "postDate": "10/14/2019 19:00:06",
      "content": "<p>For augmentations, I went with default thresholds from <code>albumentations</code> library.</p>",
      "rawMarkdown": "For augmentations, I went with default thresholds from `albumentations` library.",
      "votes": null
    },
    {
      "id": "649375",
      "postDate": "10/15/2019 09:04:22",
      "content": "<p>Ok cool, let me see the effect of this on EfficientNetb4</p>",
      "rawMarkdown": "Ok cool, let me see the effect of this on EfficientNetb4",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 648849,
      "author_name": "datachampion",
      "author_url": "",
      "post_date": "10/14/2019 17:28:21",
      "content": "<p><a href=\"/vivekwisdom\">@vivekwisdom</a>  - What threshold did you use or ranges of threshold?</p>",
      "votes": null,
      "replies": [
        {
          "id": 648904,
          "author_name": "vivekwisdom",
          "author_url": "",
          "post_date": "10/14/2019 19:00:06",
          "content": "<p>For augmentations, I went with default thresholds from <code>albumentations</code> library.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 649375,
          "author_name": "datachampion",
          "author_url": "",
          "post_date": "10/15/2019 09:04:22",
          "content": "<p>Ok cool, let me see the effect of this on EfficientNetb4</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "648245": "I see most of the public kernels have used flip and rotate augmentations. Although they work well with most of the image datasets, I feel we have a unique situation here and does require special consideration for this dataset.\n\nGiven we are dealing with **Satellite images lighting (Solarize Augmentation)** might play a key role in balancing the dataset? *I did try a couple of combinations and resulted in 0.659, so it seems worth trying.*\n\nWhat other augmentation you feel could be important in this dataset? Although I haven't tried what's difference b/w TTA vs non-TTA models? Looking for insightful discussion. TIA",
    "648849": "vivekwisdom  - What threshold did you use or ranges of threshold?",
    "648904": "For augmentations, I went with default thresholds from `albumentations` library.",
    "649375": "Ok cool, let me see the effect of this on EfficientNetb4"
  },
  "source": "meta"
}