{
  "id": 104856,
  "title": "Large Model Small Images vs Small Model Large Images",
  "url": "/competitions/recursion-cellular-image-classification/discussion/104856",
  "author_name": "",
  "post_date": "2019-08-19T16:15:01.106004400Z",
  "votes": 6,
  "comment_count": 5,
  "views": 0,
  "content": "<p>As stated in the topic, which approach would make more sense in image classification task in general (especially for experimenting or in this competition)? </p>\n\n<p>Seeing that most papers reporting results on ImageNet are using images that have been resized to smaller dimensions, does this put smaller models at an unfair disadvantage? </p>\n\n<p>Since larger models are better at producing state of the art results when using smaller image sizes, small models can handle larger images giving the same computing resources.</p>\n\n<p>Please let me know your thoughts. Thank you! </p>",
  "messages": [
    {
      "id": "602906",
      "postDate": "08/19/2019 16:15:01",
      "content": "<p>As stated in the topic, which approach would make more sense in image classification task in general (especially for experimenting or in this competition)? </p>\n\n<p>Seeing that most papers reporting results on ImageNet are using images that have been resized to smaller dimensions, does this put smaller models at an unfair disadvantage? </p>\n\n<p>Since larger models are better at producing state of the art results when using smaller image sizes, small models can handle larger images giving the same computing resources.</p>\n\n<p>Please let me know your thoughts. Thank you! </p>",
      "rawMarkdown": "As stated in the topic, which approach would make more sense in image classification task in general (especially for experimenting or in this competition)? \n\nSeeing that most papers reporting results on ImageNet are using images that have been resized to smaller dimensions, does this put smaller models at an unfair disadvantage? \n\nSince larger models are better at producing state of the art results when using smaller image sizes, small models can handle larger images giving the same computing resources.\n\nPlease let me know your thoughts. Thank you!",
      "votes": null
    },
    {
      "id": "603337",
      "postDate": "08/20/2019 06:51:39",
      "content": "<p>I haven't tried larger models yet, I'm still using ResNet18 and I reach a plateau after roughly 9 epochs (0.3 validation accuracy). What do you mean by <em>large models</em>: like DenseNet121 rather than ResNet34?</p>",
      "rawMarkdown": "I haven't tried larger models yet, I'm still using ResNet18 and I reach a plateau after roughly 9 epochs (0.3 validation accuracy). What do you mean by *large models*: like DenseNet121 rather than ResNet34?",
      "votes": null
    },
    {
      "id": "603645",
      "postDate": "08/20/2019 14:16:16",
      "content": "<p>more like DenseNet201 or SE-ResNet or Se-ResNeXt used by competitors in other competitions.</p>",
      "rawMarkdown": "more like DenseNet201 or SE-ResNet or Se-ResNeXt used by competitors in other competitions.",
      "votes": null
    },
    {
      "id": "605865",
      "postDate": "08/22/2019 21:01:48",
      "content": "<p>My findings so far: for a limited amount of GPU memory, we can either go for higher resolution images with small models or lower resolution images with large models. Apparently, we can see this as a trade-off between resolution and model depths. There is a stronger preference to using larger models compared to using higher resolution images. The former gives a significant improvement over the latter. Any input is welcome. Thank you.</p>",
      "rawMarkdown": "My findings so far: for a limited amount of GPU memory, we can either go for higher resolution images with small models or lower resolution images with large models. Apparently, we can see this as a trade-off between resolution and model depths. There is a stronger preference to using larger models compared to using higher resolution images. The former gives a significant improvement over the latter. Any input is welcome. Thank you.",
      "votes": null
    },
    {
      "id": "622132",
      "postDate": "09/09/2019 10:06:13",
      "content": "<p>Hi!\nMay I ask, do you use any augmentation techniques?\nI can't get any results similar to yours using ResNet18 :(</p>",
      "rawMarkdown": "Hi!\nMay I ask, do you use any augmentation techniques?\nI can't get any results similar to yours using ResNet18 :(",
      "votes": null
    },
    {
      "id": "622526",
      "postDate": "09/09/2019 19:11:37",
      "content": "<p>Hi, I am using only flipping and rotating. This is a score of a larger model. Good luck!</p>",
      "rawMarkdown": "Hi, I am using only flipping and rotating. This is a score of a larger model. Good luck!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 603337,
      "author_name": "lorenzofabbri92",
      "author_url": "",
      "post_date": "08/20/2019 06:51:39",
      "content": "<p>I haven't tried larger models yet, I'm still using ResNet18 and I reach a plateau after roughly 9 epochs (0.3 validation accuracy). What do you mean by <em>large models</em>: like DenseNet121 rather than ResNet34?</p>",
      "votes": null,
      "replies": [
        {
          "id": 603645,
          "author_name": "wjshenggggg",
          "author_url": "",
          "post_date": "08/20/2019 14:16:16",
          "content": "<p>more like DenseNet201 or SE-ResNet or Se-ResNeXt used by competitors in other competitions.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 622132,
          "author_name": "rafailfridman",
          "author_url": "",
          "post_date": "09/09/2019 10:06:13",
          "content": "<p>Hi!\nMay I ask, do you use any augmentation techniques?\nI can't get any results similar to yours using ResNet18 :(</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 622526,
          "author_name": "wjshenggggg",
          "author_url": "",
          "post_date": "09/09/2019 19:11:37",
          "content": "<p>Hi, I am using only flipping and rotating. This is a score of a larger model. Good luck!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 605865,
      "author_name": "wjshenggggg",
      "author_url": "",
      "post_date": "08/22/2019 21:01:48",
      "content": "<p>My findings so far: for a limited amount of GPU memory, we can either go for higher resolution images with small models or lower resolution images with large models. Apparently, we can see this as a trade-off between resolution and model depths. There is a stronger preference to using larger models compared to using higher resolution images. The former gives a significant improvement over the latter. Any input is welcome. Thank you.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "602906": "As stated in the topic, which approach would make more sense in image classification task in general (especially for experimenting or in this competition)? \n\nSeeing that most papers reporting results on ImageNet are using images that have been resized to smaller dimensions, does this put smaller models at an unfair disadvantage? \n\nSince larger models are better at producing state of the art results when using smaller image sizes, small models can handle larger images giving the same computing resources.\n\nPlease let me know your thoughts. Thank you!",
    "603337": "I haven't tried larger models yet, I'm still using ResNet18 and I reach a plateau after roughly 9 epochs (0.3 validation accuracy). What do you mean by *large models*: like DenseNet121 rather than ResNet34?",
    "603645": "more like DenseNet201 or SE-ResNet or Se-ResNeXt used by competitors in other competitions.",
    "605865": "My findings so far: for a limited amount of GPU memory, we can either go for higher resolution images with small models or lower resolution images with large models. Apparently, we can see this as a trade-off between resolution and model depths. There is a stronger preference to using larger models compared to using higher resolution images. The former gives a significant improvement over the latter. Any input is welcome. Thank you.",
    "622132": "Hi!\nMay I ask, do you use any augmentation techniques?\nI can't get any results similar to yours using ResNet18 :(",
    "622526": "Hi, I am using only flipping and rotating. This is a score of a larger model. Good luck!"
  },
  "source": "meta"
}