{
  "id": 102073,
  "title": "Fixing the train-test resolution discrepancy",
  "url": "/competitions/aptos2019-blindness-detection/discussion/102073",
  "author_name": "",
  "post_date": "2019-07-30T19:04:47.740541Z",
  "votes": 12,
  "comment_count": 5,
  "views": 0,
  "content": "<p>This looks promising!!! Hope it helps in ur experiments.</p>\n\n<p><img src=\"https://raw.githubusercontent.com/facebookresearch/FixRes/master/image/image2.png\" alt=\"Image\"></p>\n\n<h2>Abstract</h2>\n\n<p><code>\nData-augmentation is key to the training of neural networks for image classification. This paper first shows that existing augmentations induce a significant discrepancy between the typical size of the objects seen by the classifier at train and test time. We experimentally validate that, for a target test resolution, using a lower train resolution offers better classification at test time. We then propose a simple yet effective and efficient strategy to optimize the classifier performance when the train and test resolutions differ. It involves only a computationally cheap fine-tuning of the network at the test resolution. This enables training strong classifiers using small training images. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 79.8% with one trained on 224x224 image. In addition, if we use extra training data we get 82.5% with the ResNet-50 train with 224x224 images. Conversely, when training a ResNeXt-101 32x48d pre-trained in weakly-supervised fashion on 940 million public images at resolution 224x224 and further optimizing for test resolution 320x320, we obtain a test top-1 accuracy of 86.4% (top-5: 98.0%) (single-crop). To the best of our knowledge this is the highest ImageNet single-crop, top-1 and top-5 accuracy to date.\n</code></p>\n\n<p><a href=\"https://paperswithcode.com/paper/fixing-the-train-test-resolution-discrepancy\">PAPER+CoDE</a></p>",
  "messages": [
    {
      "id": "588582",
      "postDate": "07/30/2019 19:04:47",
      "content": "<p>This looks promising!!! Hope it helps in ur experiments.</p>\n\n<p><img src=\"https://raw.githubusercontent.com/facebookresearch/FixRes/master/image/image2.png\" alt=\"Image\"></p>\n\n<h2>Abstract</h2>\n\n<p><code>\nData-augmentation is key to the training of neural networks for image classification. This paper first shows that existing augmentations induce a significant discrepancy between the typical size of the objects seen by the classifier at train and test time. We experimentally validate that, for a target test resolution, using a lower train resolution offers better classification at test time. We then propose a simple yet effective and efficient strategy to optimize the classifier performance when the train and test resolutions differ. It involves only a computationally cheap fine-tuning of the network at the test resolution. This enables training strong classifiers using small training images. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 79.8% with one trained on 224x224 image. In addition, if we use extra training data we get 82.5% with the ResNet-50 train with 224x224 images. Conversely, when training a ResNeXt-101 32x48d pre-trained in weakly-supervised fashion on 940 million public images at resolution 224x224 and further optimizing for test resolution 320x320, we obtain a test top-1 accuracy of 86.4% (top-5: 98.0%) (single-crop). To the best of our knowledge this is the highest ImageNet single-crop, top-1 and top-5 accuracy to date.\n</code></p>\n\n<p><a href=\"https://paperswithcode.com/paper/fixing-the-train-test-resolution-discrepancy\">PAPER+CoDE</a></p>",
      "rawMarkdown": "This looks promising!!! Hope it helps in ur experiments.\n\n![Image](https://raw.githubusercontent.com/facebookresearch/FixRes/master/image/image2.png)\n## Abstract\n```\nData-augmentation is key to the training of neural networks for image classification. This paper first shows that existing augmentations induce a significant discrepancy between the typical size of the objects seen by the classifier at train and test time. We experimentally validate that, for a target test resolution, using a lower train resolution offers better classification at test time. We then propose a simple yet effective and efficient strategy to optimize the classifier performance when the train and test resolutions differ. It involves only a computationally cheap fine-tuning of the network at the test resolution. This enables training strong classifiers using small training images. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 79.8% with one trained on 224x224 image. In addition, if we use extra training data we get 82.5% with the ResNet-50 train with 224x224 images. Conversely, when training a ResNeXt-101 32x48d pre-trained in weakly-supervised fashion on 940 million public images at resolution 224x224 and further optimizing for test resolution 320x320, we obtain a test top-1 accuracy of 86.4% (top-5: 98.0%) (single-crop). To the best of our knowledge this is the highest ImageNet single-crop, top-1 and top-5 accuracy to date.\n```\n\n[PAPER+CoDE](https://paperswithcode.com/paper/fixing-the-train-test-resolution-discrepancy)",
      "votes": null
    },
    {
      "id": "588692",
      "postDate": "07/30/2019 23:45:13",
      "content": "<p>Nice paper</p>",
      "rawMarkdown": "Nice paper",
      "votes": null
    },
    {
      "id": "589879",
      "postDate": "08/01/2019 14:12:18",
      "content": "<p>Thanks for sharing.</p>",
      "rawMarkdown": "Thanks for sharing.",
      "votes": null
    },
    {
      "id": "590218",
      "postDate": "08/01/2019 22:13:35",
      "content": "<p>I am not seeing how this may be used in this competition. Based on my understanding, there is no significant train-test-discrepancy, no?</p>",
      "rawMarkdown": "I am not seeing how this may be used in this competition. Based on my understanding, there is no significant train-test-discrepancy, no?",
      "votes": null
    },
    {
      "id": "590318",
      "postDate": "08/02/2019 04:17:57",
      "content": "<p>I am not sure but mayb the scaling strategy could help to match the resolutions of train-test set(which are different)</p>",
      "rawMarkdown": "I am not sure but mayb the scaling strategy could help to match the resolutions of train-test set(which are different)",
      "votes": null
    },
    {
      "id": "590322",
      "postDate": "08/02/2019 04:40:19",
      "content": "<p>Great paper - thanks for sharing.</p>\n\n<p>I'm thinking that we always want to consider going to opposite way in this competition: train on larger images and test on smaller images, since the eyes in the test set seem to be significantly large than train.</p>",
      "rawMarkdown": "Great paper - thanks for sharing.\n\nI'm thinking that we always want to consider going to opposite way in this competition: train on larger images and test on smaller images, since the eyes in the test set seem to be significantly large than train.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 588692,
      "author_name": "princedemo",
      "author_url": "",
      "post_date": "07/30/2019 23:45:13",
      "content": "<p>Nice paper</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 589879,
      "author_name": "suneelpatel",
      "author_url": "",
      "post_date": "08/01/2019 14:12:18",
      "content": "<p>Thanks for sharing.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 590218,
      "author_name": "tanlikesmath",
      "author_url": "",
      "post_date": "08/01/2019 22:13:35",
      "content": "<p>I am not seeing how this may be used in this competition. Based on my understanding, there is no significant train-test-discrepancy, no?</p>",
      "votes": null,
      "replies": [
        {
          "id": 590318,
          "author_name": "bibek777",
          "author_url": "",
          "post_date": "08/02/2019 04:17:57",
          "content": "<p>I am not sure but mayb the scaling strategy could help to match the resolutions of train-test set(which are different)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 590322,
      "author_name": "lextoumbourou",
      "author_url": "",
      "post_date": "08/02/2019 04:40:19",
      "content": "<p>Great paper - thanks for sharing.</p>\n\n<p>I'm thinking that we always want to consider going to opposite way in this competition: train on larger images and test on smaller images, since the eyes in the test set seem to be significantly large than train.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "588582": "This looks promising!!! Hope it helps in ur experiments.\n\n![Image](https://raw.githubusercontent.com/facebookresearch/FixRes/master/image/image2.png)\n## Abstract\n```\nData-augmentation is key to the training of neural networks for image classification. This paper first shows that existing augmentations induce a significant discrepancy between the typical size of the objects seen by the classifier at train and test time. We experimentally validate that, for a target test resolution, using a lower train resolution offers better classification at test time. We then propose a simple yet effective and efficient strategy to optimize the classifier performance when the train and test resolutions differ. It involves only a computationally cheap fine-tuning of the network at the test resolution. This enables training strong classifiers using small training images. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 79.8% with one trained on 224x224 image. In addition, if we use extra training data we get 82.5% with the ResNet-50 train with 224x224 images. Conversely, when training a ResNeXt-101 32x48d pre-trained in weakly-supervised fashion on 940 million public images at resolution 224x224 and further optimizing for test resolution 320x320, we obtain a test top-1 accuracy of 86.4% (top-5: 98.0%) (single-crop). To the best of our knowledge this is the highest ImageNet single-crop, top-1 and top-5 accuracy to date.\n```\n\n[PAPER+CoDE](https://paperswithcode.com/paper/fixing-the-train-test-resolution-discrepancy)",
    "588692": "Nice paper",
    "589879": "Thanks for sharing.",
    "590218": "I am not seeing how this may be used in this competition. Based on my understanding, there is no significant train-test-discrepancy, no?",
    "590318": "I am not sure but mayb the scaling strategy could help to match the resolutions of train-test set(which are different)",
    "590322": "Great paper - thanks for sharing.\n\nI'm thinking that we always want to consider going to opposite way in this competition: train on larger images and test on smaller images, since the eyes in the test set seem to be significantly large than train."
  },
  "source": "meta"
}