{
  "id": 142136,
  "title": "Test Time Augmentation Question",
  "url": "/competitions/flower-classification-with-tpus/discussion/142136",
  "author_name": "",
  "post_date": "2020-04-09T04:41:59.793694800Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>What is the best way to determine the predicted label from multiple predictions? Is it better to argmax the averaged softmax probabilities or argmax each set of probabilities and take the most frequent class (or are those the same thing?)</p>",
  "messages": [
    {
      "id": "802039",
      "postDate": "04/09/2020 04:41:59",
      "content": "<p>What is the best way to determine the predicted label from multiple predictions? Is it better to argmax the averaged softmax probabilities or argmax each set of probabilities and take the most frequent class (or are those the same thing?)</p>",
      "rawMarkdown": "What is the best way to determine the predicted label from multiple predictions? Is it better to argmax the averaged softmax probabilities or argmax each set of probabilities and take the most frequent class (or are those the same thing?)",
      "votes": null
    },
    {
      "id": "807940",
      "postDate": "04/15/2020 03:55:05",
      "content": "<p>They are not the same thing. You should try both and see which yields a higher validation score. Note that you can also average the logits and then apply softmax to the average. This is a third option.</p>",
      "rawMarkdown": "They are not the same thing. You should try both and see which yields a higher validation score. Note that you can also average the logits and then apply softmax to the average. This is a third option.",
      "votes": null
    },
    {
      "id": "815885",
      "postDate": "04/21/2020 23:02:42",
      "content": "<p>Thanks for the suggestions. I took a bit of a dive into the weeds to see how to calculate each <a href=\"https://www.kaggle.com/calebeverett/comparison-of-tta-prediction-procedures\">here</a></p>",
      "rawMarkdown": "Thanks for the suggestions. I took a bit of a dive into the weeds to see how to calculate each [here](https://www.kaggle.com/calebeverett/comparison-of-tta-prediction-procedures)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 807940,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "04/15/2020 03:55:05",
      "content": "<p>They are not the same thing. You should try both and see which yields a higher validation score. Note that you can also average the logits and then apply softmax to the average. This is a third option.</p>",
      "votes": null,
      "replies": [
        {
          "id": 815885,
          "author_name": "calebeverett",
          "author_url": "",
          "post_date": "04/21/2020 23:02:42",
          "content": "<p>Thanks for the suggestions. I took a bit of a dive into the weeds to see how to calculate each <a href=\"https://www.kaggle.com/calebeverett/comparison-of-tta-prediction-procedures\">here</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "802039": "What is the best way to determine the predicted label from multiple predictions? Is it better to argmax the averaged softmax probabilities or argmax each set of probabilities and take the most frequent class (or are those the same thing?)",
    "807940": "They are not the same thing. You should try both and see which yields a higher validation score. Note that you can also average the logits and then apply softmax to the average. This is a third option.",
    "815885": "Thanks for the suggestions. I took a bit of a dive into the weeds to see how to calculate each [here](https://www.kaggle.com/calebeverett/comparison-of-tta-prediction-procedures)"
  },
  "source": "meta"
}