{
  "id": 182894,
  "title": "Are EfficientNets useful in this competition ?",
  "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/182894",
  "author_name": "Abdur Rehman",
  "post_date": "2020-09-14T17:54:08.904000",
  "votes": 5,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I am a newbie computer vision so I have a question for experts here.</p>\n<p>I have seen recently a large use of efficientnet in computer vision competitions. In Melanoma competition, winners best model was efficientnet b7 with heavy augmentations but I am surprised I could not see anyone using efficientnet in this competition. Can we use efficientnet in this competition to get good results as the previous competitions? If no, then why ?</p>",
  "messages": [
    {
      "id": 1010386,
      "postDate": "2020-09-14T17:54:08.903Z",
      "content": "<p>I am a newbie computer vision so I have a question for experts here.</p>\n<p>I have seen recently a large use of efficientnet in computer vision competitions. In Melanoma competition, winners best model was efficientnet b7 with heavy augmentations but I am surprised I could not see anyone using efficientnet in this competition. Can we use efficientnet in this competition to get good results as the previous competitions? If no, then why ?</p>",
      "rawMarkdown": "I am a newbie computer vision so I have a question for experts here.\n\nI have seen recently a large use of efficientnet in computer vision competitions. In Melanoma competition, winners best model was efficientnet b7 with heavy augmentations but I am surprised I could not see anyone using efficientnet in this competition. Can we use efficientnet in this competition to get good results as the previous competitions? If no, then why ?",
      "votes": 5
    },
    {
      "id": 1010422,
      "postDate": "2020-09-14T18:20:31.593Z",
      "content": "<p>I'm not sure if you need internet access to load efficientnet. I haven't tested, but efficientnet is not in the standard Tensorflow library (I think). You could probably store it in a dataset and access it in your committed notebook.</p>\n<p>Rather than explore this, I'm using other popular models, just to get the basic framework running.</p>\n<p>Easy enough to switch once I'm ready for tuning.</p>\n<p>-Rich</p>",
      "rawMarkdown": "I'm not sure if you need internet access to load efficientnet. I haven't tested, but efficientnet is not in the standard Tensorflow library (I think). You could probably store it in a dataset and access it in your committed notebook.\n\nRather than explore this, I'm using other popular models, just to get the basic framework running.\n\nEasy enough to switch once I'm ready for tuning.\n\n-Rich",
      "votes": 1,
      "replies": [
        {
          "id": 1010449,
          "postDate": "2020-09-14T18:44:18.307Z",
          "content": "<p>Thanks for the reply.</p>\n<p>Could you plz mention which models are you using right now for experiments and they don't need internet access? I think we can't use the internet for inference kernel but we can use it in training kernel, correct me if I am wrong.</p>",
          "rawMarkdown": "Thanks for the reply.\n\nCould you plz mention which models are you using right now for experiments and they don't need internet access? I think we can't use the internet for inference kernel but we can use it in training kernel, correct me if I am wrong.",
          "replies": [
            {
              "id": 1010452,
              "postDate": "2020-09-14T18:52:13.607Z",
              "content": "<p>I've successfully used Resnet101, DenseNet and Xception.</p>\n<p>I don't have any information which one is best. I'm continually adding metadata and training on larger datasets, so I have no comparable results. Not even sure the results are better than guessing.</p>\n<p>When I tried to load the Efficientnet model and weights without Internet, I got an error. I think something in EfficientNet doesn't get saved correctly with model.save. In the past, I just rebuilt the model, and then loaded the saved weights. But I didn't attempt to fix this problem for this contest.</p>\n<p>As I said, I'm far from tuning a model. Just trying to get something that starts working.</p>\n<p>-Rich</p>",
              "rawMarkdown": "I've successfully used Resnet101, DenseNet and Xception.\n\nI don't have any information which one is best. I'm continually adding metadata and training on larger datasets, so I have no comparable results. Not even sure the results are better than guessing.\n\nWhen I tried to load the Efficientnet model and weights without Internet, I got an error. I think something in EfficientNet doesn't get saved correctly with model.save. In the past, I just rebuilt the model, and then loaded the saved weights. But I didn't attempt to fix this problem for this contest.\n\nAs I said, I'm far from tuning a model. Just trying to get something that starts working.\n\n-Rich",
              "votes": 4
            },
            {
              "id": 1010456,
              "postDate": "2020-09-14T18:58:03.337Z",
              "content": "<p>Thanks for sharing the approach.</p>",
              "rawMarkdown": "Thanks for sharing the approach."
            }
          ]
        },
        {
          "id": 1013009,
          "postDate": "2020-09-16T13:06:51.787Z",
          "content": "<p>I guess you can train in another notebook and use the output (model) from there in your inference notebook. The competition specified that no internet access is allowed, but notebook outputs can be linked i believe. Otherwise, you can just upload your model as an input.</p>",
          "rawMarkdown": "I guess you can train in another notebook and use the output (model) from there in your inference notebook. The competition specified that no internet access is allowed, but notebook outputs can be linked i believe. Otherwise, you can just upload your model as an input."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1010422,
      "author_name": "quadcore/Richard Epstein",
      "author_url": "",
      "post_date": "2020-09-14T18:20:31.593000",
      "content": "<p>I'm not sure if you need internet access to load efficientnet. I haven't tested, but efficientnet is not in the standard Tensorflow library (I think). You could probably store it in a dataset and access it in your committed notebook.</p>\n<p>Rather than explore this, I'm using other popular models, just to get the basic framework running.</p>\n<p>Easy enough to switch once I'm ready for tuning.</p>\n<p>-Rich</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1010449,
          "author_name": "Abdur Rehman",
          "author_url": "",
          "post_date": "2020-09-14T18:44:18.307000",
          "content": "<p>Thanks for the reply.</p>\n<p>Could you plz mention which models are you using right now for experiments and they don't need internet access? I think we can't use the internet for inference kernel but we can use it in training kernel, correct me if I am wrong.</p>",
          "votes": 0,
          "replies": [
            {
              "id": 1010452,
              "author_name": "quadcore/Richard Epstein",
              "author_url": "",
              "post_date": "2020-09-14T18:52:13.607000",
              "content": "<p>I've successfully used Resnet101, DenseNet and Xception.</p>\n<p>I don't have any information which one is best. I'm continually adding metadata and training on larger datasets, so I have no comparable results. Not even sure the results are better than guessing.</p>\n<p>When I tried to load the Efficientnet model and weights without Internet, I got an error. I think something in EfficientNet doesn't get saved correctly with model.save. In the past, I just rebuilt the model, and then loaded the saved weights. But I didn't attempt to fix this problem for this contest.</p>\n<p>As I said, I'm far from tuning a model. Just trying to get something that starts working.</p>\n<p>-Rich</p>",
              "votes": 4,
              "replies": []
            },
            {
              "id": 1010456,
              "author_name": "Abdur Rehman",
              "author_url": "",
              "post_date": "2020-09-14T18:58:03.337000",
              "content": "<p>Thanks for sharing the approach.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 1013009,
          "author_name": "Jesudas DSouza",
          "author_url": "",
          "post_date": "2020-09-16T13:06:51.787000",
          "content": "<p>I guess you can train in another notebook and use the output (model) from there in your inference notebook. The competition specified that no internet access is allowed, but notebook outputs can be linked i believe. Otherwise, you can just upload your model as an input.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1010386": "I am a newbie computer vision so I have a question for experts here.\n\nI have seen recently a large use of efficientnet in computer vision competitions. In Melanoma competition, winners best model was efficientnet b7 with heavy augmentations but I am surprised I could not see anyone using efficientnet in this competition. Can we use efficientnet in this competition to get good results as the previous competitions? If no, then why ?",
    "1010422": "I'm not sure if you need internet access to load efficientnet. I haven't tested, but efficientnet is not in the standard Tensorflow library (I think). You could probably store it in a dataset and access it in your committed notebook.\n\nRather than explore this, I'm using other popular models, just to get the basic framework running.\n\nEasy enough to switch once I'm ready for tuning.\n\n-Rich"
  }
}