{
  "id": 106609,
  "title": "TTA performance in this competition",
  "url": "/competitions/aptos2019-blindness-detection/discussion/106609",
  "author_name": "",
  "post_date": "2019-08-30T08:24:10.579524900Z",
  "votes": 3,
  "comment_count": 9,
  "views": 0,
  "content": "<p>In our case we did not get any LB improvement after using TTA. The reason for this might be not adding right set of augmentations. If anyone can suggest right set of augmentations or whether TTA is worth looking into would really help us. </p>",
  "messages": [
    {
      "id": "613118",
      "postDate": "08/30/2019 08:24:10",
      "content": "<p>In our case we did not get any LB improvement after using TTA. The reason for this might be not adding right set of augmentations. If anyone can suggest right set of augmentations or whether TTA is worth looking into would really help us. </p>",
      "rawMarkdown": "In our case we did not get any LB improvement after using TTA. The reason for this might be not adding right set of augmentations. If anyone can suggest right set of augmentations or whether TTA is worth looking into would really help us.",
      "votes": null
    },
    {
      "id": "613235",
      "postDate": "08/30/2019 10:37:16",
      "content": "<p>RandomHorizontalFlips + RandomRotate90 give me some points. But I really doubt that you didn`t try it)</p>",
      "rawMarkdown": "RandomHorizontalFlips + RandomRotate90 give me some points. But I really doubt that you didn`t try it)",
      "votes": null
    },
    {
      "id": "613375",
      "postDate": "08/30/2019 13:05:47",
      "content": "<p>We tried TTA but did not get any LB improvement. Our guess is Random Rotation that we used was the problem in the augmentation sets.</p>",
      "rawMarkdown": "We tried TTA but did not get any LB improvement. Our guess is Random Rotation that we used was the problem in the augmentation sets.",
      "votes": null
    },
    {
      "id": "613761",
      "postDate": "08/30/2019 19:45:31",
      "content": "<p>We have an update regarding TTA. We used TTA (flip_lr, flip_ud) on our single model and the score improved from 0.819 to 0.824. We think that random rotation should not be used in TTA.</p>",
      "rawMarkdown": "We have an update regarding TTA. We used TTA (flip_lr, flip_ud) on our single model and the score improved from 0.819 to 0.824. We think that random rotation should not be used in TTA.",
      "votes": null
    },
    {
      "id": "613892",
      "postDate": "08/30/2019 22:54:28",
      "content": "<p>I tried several types of TTA, only d4 did ~0.001 better, other types were ~0.005 worse</p>",
      "rawMarkdown": "I tried several types of TTA, only d4 did ~0.001 better, other types were ~0.005 worse",
      "votes": null
    },
    {
      "id": "614096",
      "postDate": "08/31/2019 04:23:19",
      "content": "<p>Hmm, Mb I should try just flips. I will report results later)</p>",
      "rawMarkdown": "Hmm, Mb I should try just flips. I will report results later)",
      "votes": null
    },
    {
      "id": "614209",
      "postDate": "08/31/2019 08:04:47",
      "content": "<p>If we plan to use TTA, should we use those same augmentations on our validation set while training? Or regardless of TTA should the same augmentations used on the training set be applied to the validation set?</p>",
      "rawMarkdown": "If we plan to use TTA, should we use those same augmentations on our validation set while training? Or regardless of TTA should the same augmentations used on the training set be applied to the validation set?",
      "votes": null
    },
    {
      "id": "614213",
      "postDate": "08/31/2019 08:08:45",
      "content": "<p>Tried only flips on my best model, lost 0.003 </p>",
      "rawMarkdown": "Tried only flips on my best model, lost 0.003",
      "votes": null
    },
    {
      "id": "614631",
      "postDate": "08/31/2019 19:54:17",
      "content": "<p><a href=\"/convexproblems\">@convexproblems</a> at first we added the same augmentations used in the TTA set as we used while training our model. But it did not boost our score. While the one stated above worked for us.</p>",
      "rawMarkdown": "convexproblems at first we added the same augmentations used in the TTA set as we used while training our model. But it did not boost our score. While the one stated above worked for us.",
      "votes": null
    },
    {
      "id": "614634",
      "postDate": "08/31/2019 19:56:53",
      "content": "<p><a href=\"/kupchanski\">@kupchanski</a> Our model was trained using regression loss and we used random flips augmentations while training. So, there are other factors to consider as well. </p>",
      "rawMarkdown": "kupchanski Our model was trained using regression loss and we used random flips augmentations while training. So, there are other factors to consider as well.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 613235,
      "author_name": "kupchanski",
      "author_url": "",
      "post_date": "08/30/2019 10:37:16",
      "content": "<p>RandomHorizontalFlips + RandomRotate90 give me some points. But I really doubt that you didn`t try it)</p>",
      "votes": null,
      "replies": [
        {
          "id": 613375,
          "author_name": "sabbiracoustic1006",
          "author_url": "",
          "post_date": "08/30/2019 13:05:47",
          "content": "<p>We tried TTA but did not get any LB improvement. Our guess is Random Rotation that we used was the problem in the augmentation sets.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 613761,
          "author_name": "sabbiracoustic1006",
          "author_url": "",
          "post_date": "08/30/2019 19:45:31",
          "content": "<p>We have an update regarding TTA. We used TTA (flip_lr, flip_ud) on our single model and the score improved from 0.819 to 0.824. We think that random rotation should not be used in TTA.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 614096,
          "author_name": "kupchanski",
          "author_url": "",
          "post_date": "08/31/2019 04:23:19",
          "content": "<p>Hmm, Mb I should try just flips. I will report results later)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 614209,
          "author_name": "convexproblems",
          "author_url": "",
          "post_date": "08/31/2019 08:04:47",
          "content": "<p>If we plan to use TTA, should we use those same augmentations on our validation set while training? Or regardless of TTA should the same augmentations used on the training set be applied to the validation set?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 614213,
          "author_name": "kupchanski",
          "author_url": "",
          "post_date": "08/31/2019 08:08:45",
          "content": "<p>Tried only flips on my best model, lost 0.003 </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 614631,
          "author_name": "sabbiracoustic1006",
          "author_url": "",
          "post_date": "08/31/2019 19:54:17",
          "content": "<p><a href=\"/convexproblems\">@convexproblems</a> at first we added the same augmentations used in the TTA set as we used while training our model. But it did not boost our score. While the one stated above worked for us.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 614634,
          "author_name": "sabbiracoustic1006",
          "author_url": "",
          "post_date": "08/31/2019 19:56:53",
          "content": "<p><a href=\"/kupchanski\">@kupchanski</a> Our model was trained using regression loss and we used random flips augmentations while training. So, there are other factors to consider as well. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 613892,
      "author_name": "spsancti",
      "author_url": "",
      "post_date": "08/30/2019 22:54:28",
      "content": "<p>I tried several types of TTA, only d4 did ~0.001 better, other types were ~0.005 worse</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "613118": "In our case we did not get any LB improvement after using TTA. The reason for this might be not adding right set of augmentations. If anyone can suggest right set of augmentations or whether TTA is worth looking into would really help us.",
    "613235": "RandomHorizontalFlips + RandomRotate90 give me some points. But I really doubt that you didn`t try it)",
    "613375": "We tried TTA but did not get any LB improvement. Our guess is Random Rotation that we used was the problem in the augmentation sets.",
    "613761": "We have an update regarding TTA. We used TTA (flip_lr, flip_ud) on our single model and the score improved from 0.819 to 0.824. We think that random rotation should not be used in TTA.",
    "613892": "I tried several types of TTA, only d4 did ~0.001 better, other types were ~0.005 worse",
    "614096": "Hmm, Mb I should try just flips. I will report results later)",
    "614209": "If we plan to use TTA, should we use those same augmentations on our validation set while training? Or regardless of TTA should the same augmentations used on the training set be applied to the validation set?",
    "614213": "Tried only flips on my best model, lost 0.003",
    "614631": "convexproblems at first we added the same augmentations used in the TTA set as we used while training our model. But it did not boost our score. While the one stated above worked for us.",
    "614634": "kupchanski Our model was trained using regression loss and we used random flips augmentations while training. So, there are other factors to consider as well."
  },
  "source": "meta"
}