{
  "id": 232051,
  "title": "sota: from vgg to resnet to efficient-net, then back to resnet, back to vgg again",
  "url": "/competitions/bms-molecular-translation/discussion/232051",
  "author_name": "",
  "post_date": "2021-04-11T23:54:21.168013500Z",
  "votes": 15,
  "comment_count": 2,
  "views": 0,
  "content": "<p><a href=\"https://github.com/DingXiaoH/RepVGG\" target=\"_blank\">https://github.com/DingXiaoH/RepVGG</a><br>\n<a href=\"https://arxiv.org/pdf/2101.03697.pdf\" target=\"_blank\">https://arxiv.org/pdf/2101.03697.pdf</a></p>\n<p>inference: VGG without residual connect<br>\ntraining: like resnet</p>\n<p>\"On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet.\"</p>",
  "messages": [
    {
      "id": "1270733",
      "postDate": "04/11/2021 23:54:21",
      "content": "<p><a href=\"https://github.com/DingXiaoH/RepVGG\" target=\"_blank\">https://github.com/DingXiaoH/RepVGG</a><br>\n<a href=\"https://arxiv.org/pdf/2101.03697.pdf\" target=\"_blank\">https://arxiv.org/pdf/2101.03697.pdf</a></p>\n<p>inference: VGG without residual connect<br>\ntraining: like resnet</p>\n<p>\"On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet.\"</p>",
      "rawMarkdown": "https://github.com/DingXiaoH/RepVGG\nhttps://arxiv.org/pdf/2101.03697.pdf\n\ninference: VGG without residual connect\ntraining: like resnet\n\n\"On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet.\"",
      "votes": null
    },
    {
      "id": "1270887",
      "postDate": "04/12/2021 06:16:21",
      "content": "<p>Repvggs have not performed as well as I’d hoped here so far. </p>",
      "rawMarkdown": "Repvggs have not performed as well as I’d hoped here so far.",
      "votes": null
    },
    {
      "id": "1270918",
      "postDate": "04/12/2021 06:43:38",
      "content": "<p>Unfortunately, unlike the theoretical flops, group convolution is very slow, but if we set cudnn.benchmark = True, we can training quickly!</p>",
      "rawMarkdown": "Unfortunately, unlike the theoretical flops, group convolution is very slow, but if we set cudnn.benchmark = True, we can training quickly!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1270887,
      "author_name": "interneuron",
      "author_url": "",
      "post_date": "04/12/2021 06:16:21",
      "content": "<p>Repvggs have not performed as well as I’d hoped here so far. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1270918,
      "author_name": "gwanghan",
      "author_url": "",
      "post_date": "04/12/2021 06:43:38",
      "content": "<p>Unfortunately, unlike the theoretical flops, group convolution is very slow, but if we set cudnn.benchmark = True, we can training quickly!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1270733": "https://github.com/DingXiaoH/RepVGG\nhttps://arxiv.org/pdf/2101.03697.pdf\n\ninference: VGG without residual connect\ntraining: like resnet\n\n\"On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet.\"",
    "1270887": "Repvggs have not performed as well as I’d hoped here so far.",
    "1270918": "Unfortunately, unlike the theoretical flops, group convolution is very slow, but if we set cudnn.benchmark = True, we can training quickly!"
  },
  "source": "meta"
}