{
  "id": 156482,
  "title": "Beginner Advice",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/156482",
  "author_name": "",
  "post_date": "2020-06-06T10:40:19.414227500Z",
  "votes": null,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi I am Beginner in Computer Vision. I performed several augmentations and trained with densenet with imagesize 450 and only got AUC around 0.62. Since probability rank and AUC is used as a metric, I can't find any steps to improve my model. Any suggestions and ideas will be welcomed. Thanks in advance.</p>",
  "messages": [
    {
      "id": "875984",
      "postDate": "06/06/2020 10:40:19",
      "content": "<p>Hi I am Beginner in Computer Vision. I performed several augmentations and trained with densenet with imagesize 450 and only got AUC around 0.62. Since probability rank and AUC is used as a metric, I can't find any steps to improve my model. Any suggestions and ideas will be welcomed. Thanks in advance.</p>",
      "rawMarkdown": "Hi I am Beginner in Computer Vision. I performed several augmentations and trained with densenet with imagesize 450 and only got AUC around 0.62. Since probability rank and AUC is used as a metric, I can't find any steps to improve my model. Any suggestions and ideas will be welcomed. Thanks in advance.",
      "votes": null
    },
    {
      "id": "875994",
      "postDate": "06/06/2020 10:53:24",
      "content": "<p>!pip install -q efficientnet</p>\n\n<p>import efficientnet.tfkeras as efn</p>\n\n<p>EPOCHS = 50</p>\n\n<p>eg IMAGE_SIZE = [512, 512]</p>\n\n<p>model = tf.keras.Sequential([\n        efn.EfficientNetB5(\n            input_shape=(None,None,3), #(*IMAGE_SIZE, 3),\n            weights='noisy-student',\n            include_top=False\n        ),\n        L.GlobalAveragePooling2D(),\n        L.Dense(1, activation='sigmoid',dtype='float32')\n    ])</p>\n\n<p>use loss with label_smoothing or focal loss and learning rate scheduler with a warmup</p>\n\n<p>happy training</p>",
      "rawMarkdown": "!pip install -q efficientnet\n\nimport efficientnet.tfkeras as efn\n\nEPOCHS = 50\n\neg IMAGE_SIZE = [512, 512]\n\nmodel = tf.keras.Sequential([\n        efn.EfficientNetB5(\n            input_shape=(None,None,3), #(*IMAGE_SIZE, 3),\n            weights='noisy-student',\n            include_top=False\n        ),\n        L.GlobalAveragePooling2D(),\n        L.Dense(1, activation='sigmoid',dtype='float32')\n    ])\n\nuse loss with label_smoothing or focal loss and learning rate scheduler with a warmup\n\nhappy training",
      "votes": null
    },
    {
      "id": "876389",
      "postDate": "06/06/2020 16:47:59",
      "content": "<p>Great!</p>",
      "rawMarkdown": "Great!",
      "votes": null
    },
    {
      "id": "876896",
      "postDate": "06/07/2020 06:22:55",
      "content": "<ul>\n<li>For starters, visually check the images you are feeding to your densenet and check the outputs that are coming out. </li>\n<li>First try a simple Resnet 18 or 34 with 224x224 image size, few epochs and no augmentations without external data. It should get you atleast 0.8 AUC.</li>\n<li>Remember to handle the class imbalance.</li>\n<li><a href=\"http://karpathy.github.io/2019/04/25/recipe/\">Neural net training fails silently.</a></li>\n</ul>",
      "rawMarkdown": "For starters, visually check the images you are feeding to your densenet and check the outputs that are coming out. \n- First try a simple Resnet 18 or 34 with 224x224 image size, few epochs and no augmentations without external data. It should get you atleast 0.8 AUC.\n- Remember to handle the class imbalance.\n- [Neural net training fails silently.](http://karpathy.github.io/2019/04/25/recipe/)",
      "votes": null
    },
    {
      "id": "877070",
      "postDate": "06/07/2020 09:42:09",
      "content": "<p>Thanks <a href=\"/romanweilguny\">@romanweilguny</a>. I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel <a href=\"https://www.kaggle.com/aakashveera/dataset-creater-siim/\">https://www.kaggle.com/aakashveera/dataset-creater-siim/</a></p>",
      "rawMarkdown": "Thanks @romanweilguny. I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel [https://www.kaggle.com/aakashveera/dataset-creater-siim/](https://www.kaggle.com/aakashveera/dataset-creater-siim/)",
      "votes": null
    },
    {
      "id": "877074",
      "postDate": "06/07/2020 09:49:02",
      "content": "<p>I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel <a href=\"https://www.kaggle.com/aakashveera/dataset-creater-siim/\">https://www.kaggle.com/aakashveera/dataset-creater-siim/</a></p>",
      "rawMarkdown": "I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel [https://www.kaggle.com/aakashveera/dataset-creater-siim/](https://www.kaggle.com/aakashveera/dataset-creater-siim/)",
      "votes": null
    },
    {
      "id": "877150",
      "postDate": "06/07/2020 11:12:01",
      "content": "<blockquote>\n  <p>Notebook Output File Datasets\n  Creating a dataset from a Notebook’s output files will let you create reproducible data pipelines. To \n  create a dataset from a Notebook’s output files, click on the icon in the uploader and search for your &gt; Notebook. Alternatively, you can click “Create Dataset” from the Output tab on your rendered \n  Notebook. Then, select the files you want to use in your dataset.</p>\n</blockquote>\n\n<p><a href=\"https://www.kaggle.com/docs/datasets#creating-datasets-from-various-connectors\">https://www.kaggle.com/docs/datasets#creating-datasets-from-various-connectors</a></p>",
      "rawMarkdown": "&gt; Notebook Output File Datasets\n&gt; Creating a dataset from a Notebook’s output files will let you create reproducible data pipelines. To \n&gt; create a dataset from a Notebook’s output files, click on the icon in the uploader and search for your &gt; Notebook. Alternatively, you can click “Create Dataset” from the Output tab on your rendered \n&gt; Notebook. Then, select the files you want to use in your dataset.\n\nhttps://www.kaggle.com/docs/datasets#creating-datasets-from-various-connectors",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 875994,
      "author_name": "romanweilguny",
      "author_url": "",
      "post_date": "06/06/2020 10:53:24",
      "content": "<p>!pip install -q efficientnet</p>\n\n<p>import efficientnet.tfkeras as efn</p>\n\n<p>EPOCHS = 50</p>\n\n<p>eg IMAGE_SIZE = [512, 512]</p>\n\n<p>model = tf.keras.Sequential([\n        efn.EfficientNetB5(\n            input_shape=(None,None,3), #(*IMAGE_SIZE, 3),\n            weights='noisy-student',\n            include_top=False\n        ),\n        L.GlobalAveragePooling2D(),\n        L.Dense(1, activation='sigmoid',dtype='float32')\n    ])</p>\n\n<p>use loss with label_smoothing or focal loss and learning rate scheduler with a warmup</p>\n\n<p>happy training</p>",
      "votes": null,
      "replies": [
        {
          "id": 876389,
          "author_name": "adakoda",
          "author_url": "",
          "post_date": "06/06/2020 16:47:59",
          "content": "<p>Great!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 877070,
          "author_name": "aakashveera",
          "author_url": "",
          "post_date": "06/07/2020 09:42:09",
          "content": "<p>Thanks <a href=\"/romanweilguny\">@romanweilguny</a>. I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel <a href=\"https://www.kaggle.com/aakashveera/dataset-creater-siim/\">https://www.kaggle.com/aakashveera/dataset-creater-siim/</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 876896,
      "author_name": "utsavnandi",
      "author_url": "",
      "post_date": "06/07/2020 06:22:55",
      "content": "<ul>\n<li>For starters, visually check the images you are feeding to your densenet and check the outputs that are coming out. </li>\n<li>First try a simple Resnet 18 or 34 with 224x224 image size, few epochs and no augmentations without external data. It should get you atleast 0.8 AUC.</li>\n<li>Remember to handle the class imbalance.</li>\n<li><a href=\"http://karpathy.github.io/2019/04/25/recipe/\">Neural net training fails silently.</a></li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 877074,
          "author_name": "aakashveera",
          "author_url": "",
          "post_date": "06/07/2020 09:49:02",
          "content": "<p>I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel <a href=\"https://www.kaggle.com/aakashveera/dataset-creater-siim/\">https://www.kaggle.com/aakashveera/dataset-creater-siim/</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 877150,
          "author_name": "utsavnandi",
          "author_url": "",
          "post_date": "06/07/2020 11:12:01",
          "content": "<blockquote>\n  <p>Notebook Output File Datasets\n  Creating a dataset from a Notebook’s output files will let you create reproducible data pipelines. To \n  create a dataset from a Notebook’s output files, click on the icon in the uploader and search for your &gt; Notebook. Alternatively, you can click “Create Dataset” from the Output tab on your rendered \n  Notebook. Then, select the files you want to use in your dataset.</p>\n</blockquote>\n\n<p><a href=\"https://www.kaggle.com/docs/datasets#creating-datasets-from-various-connectors\">https://www.kaggle.com/docs/datasets#creating-datasets-from-various-connectors</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "875984": "Hi I am Beginner in Computer Vision. I performed several augmentations and trained with densenet with imagesize 450 and only got AUC around 0.62. Since probability rank and AUC is used as a metric, I can't find any steps to improve my model. Any suggestions and ideas will be welcomed. Thanks in advance.",
    "875994": "!pip install -q efficientnet\n\nimport efficientnet.tfkeras as efn\n\nEPOCHS = 50\n\neg IMAGE_SIZE = [512, 512]\n\nmodel = tf.keras.Sequential([\n        efn.EfficientNetB5(\n            input_shape=(None,None,3), #(*IMAGE_SIZE, 3),\n            weights='noisy-student',\n            include_top=False\n        ),\n        L.GlobalAveragePooling2D(),\n        L.Dense(1, activation='sigmoid',dtype='float32')\n    ])\n\nuse loss with label_smoothing or focal loss and learning rate scheduler with a warmup\n\nhappy training",
    "876389": "Great!",
    "876896": "For starters, visually check the images you are feeding to your densenet and check the outputs that are coming out. \n- First try a simple Resnet 18 or 34 with 224x224 image size, few epochs and no augmentations without external data. It should get you atleast 0.8 AUC.\n- Remember to handle the class imbalance.\n- [Neural net training fails silently.](http://karpathy.github.io/2019/04/25/recipe/)",
    "877070": "Thanks @romanweilguny. I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel [https://www.kaggle.com/aakashveera/dataset-creater-siim/](https://www.kaggle.com/aakashveera/dataset-creater-siim/)",
    "877074": "I created Image augments from the external dataset and saved it under kaggle/working and then hit commit. Now the kernel shows the output visualizations but there is no link to download/add new dataset. What is the problem? Link to the kernel [https://www.kaggle.com/aakashveera/dataset-creater-siim/](https://www.kaggle.com/aakashveera/dataset-creater-siim/)",
    "877150": "&gt; Notebook Output File Datasets\n&gt; Creating a dataset from a Notebook’s output files will let you create reproducible data pipelines. To \n&gt; create a dataset from a Notebook’s output files, click on the icon in the uploader and search for your &gt; Notebook. Alternatively, you can click “Create Dataset” from the Output tab on your rendered \n&gt; Notebook. Then, select the files you want to use in your dataset.\n\nhttps://www.kaggle.com/docs/datasets#creating-datasets-from-various-connectors"
  },
  "source": "meta"
}