{
  "id": 217340,
  "title": "Microsoft Vision Model ResNet50 - PyTorch Code",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/217340",
  "author_name": "",
  "post_date": "2021-02-06T12:22:52.033148400Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>This pretrained model was released a few days ago:<br>\n<a href=\"https://www.microsoft.com/en-us/research/blog/microsoft-vision-model-resnet-50-combines-web-scale-data-and-multi-task-learning-to-achieve-state-of-the-art/\" target=\"_blank\">https://www.microsoft.com/en-us/research/blog/microsoft-vision-model-resnet-50-combines-web-scale-data-and-multi-task-learning-to-achieve-state-of-the-art/</a><br>\n\"Microsoft Vision Model ResNet-50 is a large pretrained vision model created by the Multimedia Group at Microsoft Bing. The model is built using the search engine’s web-scale image data in order to power its Image Search and Visual Search.\"</p>\n<p>They report much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google Big Transfer.    </p>",
  "messages": [
    {
      "id": "1188653",
      "postDate": "02/06/2021 12:22:52",
      "content": "<p>This pretrained model was released a few days ago:<br>\n<a href=\"https://www.microsoft.com/en-us/research/blog/microsoft-vision-model-resnet-50-combines-web-scale-data-and-multi-task-learning-to-achieve-state-of-the-art/\" target=\"_blank\">https://www.microsoft.com/en-us/research/blog/microsoft-vision-model-resnet-50-combines-web-scale-data-and-multi-task-learning-to-achieve-state-of-the-art/</a><br>\n\"Microsoft Vision Model ResNet-50 is a large pretrained vision model created by the Multimedia Group at Microsoft Bing. The model is built using the search engine’s web-scale image data in order to power its Image Search and Visual Search.\"</p>\n<p>They report much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google Big Transfer.    </p>",
      "rawMarkdown": "This pretrained model was released a few days ago:\nhttps://www.microsoft.com/en-us/research/blog/microsoft-vision-model-resnet-50-combines-web-scale-data-and-multi-task-learning-to-achieve-state-of-the-art/\n\"Microsoft Vision Model ResNet-50 is a large pretrained vision model created by the Multimedia Group at Microsoft Bing. The model is built using the search engine’s web-scale image data in order to power its Image Search and Visual Search.\"\n\nThey report much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google Big Transfer.",
      "votes": null
    },
    {
      "id": "1188666",
      "postDate": "02/06/2021 12:40:30",
      "content": "<blockquote>\n  <p>It has much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google<br>\n  Big Transfer.</p>\n</blockquote>\n<p>I tried few days ago and gave me poor accuracy than  even standard ResNet50.  It would be great if they released the Pytorch code (not just the package) so we can see what did they modified upon classical Resnet. </p>\n<p>The other weird thing is the model is very similar in approach with Big Transfer, at least in the blog description. But the pretrained weight is much lighter and  it requires far less resources to fine-tune than the restnet50 model in Big Transfer. </p>",
      "rawMarkdown": "> It has much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google\nBig Transfer.\n\nI tried few days ago and gave me poor accuracy than  even standard ResNet50.  It would be great if they released the Pytorch code (not just the package) so we can see what did they modified upon classical Resnet. \n\n\nThe other weird thing is the model is very similar in approach with Big Transfer, at least in the blog description. But the pretrained weight is much lighter and  it requires far less resources to fine-tune than the restnet50 model in Big Transfer.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1188666,
      "author_name": "serigne",
      "author_url": "",
      "post_date": "02/06/2021 12:40:30",
      "content": "<blockquote>\n  <p>It has much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google<br>\n  Big Transfer.</p>\n</blockquote>\n<p>I tried few days ago and gave me poor accuracy than  even standard ResNet50.  It would be great if they released the Pytorch code (not just the package) so we can see what did they modified upon classical Resnet. </p>\n<p>The other weird thing is the model is very similar in approach with Big Transfer, at least in the blog description. But the pretrained weight is much lighter and  it requires far less resources to fine-tune than the restnet50 model in Big Transfer. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1188653": "This pretrained model was released a few days ago:\nhttps://www.microsoft.com/en-us/research/blog/microsoft-vision-model-resnet-50-combines-web-scale-data-and-multi-task-learning-to-achieve-state-of-the-art/\n\"Microsoft Vision Model ResNet-50 is a large pretrained vision model created by the Multimedia Group at Microsoft Bing. The model is built using the search engine’s web-scale image data in order to power its Image Search and Visual Search.\"\n\nThey report much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google Big Transfer.",
    "1188666": "> It has much better accuracy than the standard ResNet50, and also beats OpenAI CLIP and Google\nBig Transfer.\n\nI tried few days ago and gave me poor accuracy than  even standard ResNet50.  It would be great if they released the Pytorch code (not just the package) so we can see what did they modified upon classical Resnet. \n\n\nThe other weird thing is the model is very similar in approach with Big Transfer, at least in the blog description. But the pretrained weight is much lighter and  it requires far less resources to fine-tune than the restnet50 model in Big Transfer."
  },
  "source": "meta"
}