{
  "id": 82812,
  "title": "You Should Read this notebook if you want to train model faster on TPU",
  "url": "/competitions/histopathologic-cancer-detection/discussion/82812",
  "author_name": "Umberto",
  "post_date": "2019-03-04T13:56:12.571000",
  "votes": 19,
  "comment_count": 13,
  "views": 0,
  "content": "<p>Colab jupyter notebook on how to train a ResNet50/ResNeXt50 implemented on Keras with TPU on Google colab for free. \n<a href=\"https://tinyurl.com/y54y2xs9\">Check it out!</a></p>",
  "messages": [
    {
      "id": 483355,
      "postDate": "2019-03-04T13:56:12.570Z",
      "content": "<p>Colab jupyter notebook on how to train a ResNet50/ResNeXt50 implemented on Keras with TPU on Google colab for free. \n<a href=\"https://tinyurl.com/y54y2xs9\">Check it out!</a></p>",
      "rawMarkdown": "Colab jupyter notebook on how to train a ResNet50/ResNeXt50 implemented on Keras with TPU on Google colab for free. \n[Check it out!]( https://tinyurl.com/y54y2xs9)",
      "votes": 19
    },
    {
      "id": 483472,
      "postDate": "2019-03-04T16:49:39.207Z",
      "content": "<p>Thanks for sharing Umberto!\nHow do you keep Colab from disconnecting due to user inactivity?</p>",
      "rawMarkdown": "Thanks for sharing Umberto!\nHow do you keep Colab from disconnecting due to user inactivity?",
      "votes": 1
    },
    {
      "id": 487865,
      "postDate": "2019-03-11T15:11:36.347Z",
      "content": "<p>Using which cnn model have u got 0.9757 score?</p>",
      "rawMarkdown": "Using which cnn model have u got 0.9757 score?",
      "replies": [
        {
          "id": 488926,
          "postDate": "2019-03-13T08:14:11.313Z",
          "content": "<p><a href=\"/aaryapatel\">@aaryapatel</a> I used different models and approaches in order to got 0.9757 score. I used an ensemble of  ResNet50, ResNext50, Densenet121, DenseNet169 and LGBM (VGG16 as feature extractor).</p>",
          "rawMarkdown": "@aaryapatel I used different models and approaches in order to got 0.9757 score. I used an ensemble of  ResNet50, ResNext50, Densenet121, DenseNet169 and LGBM (VGG16 as feature extractor)."
        },
        {
          "id": 490253,
          "postDate": "2019-03-14T12:37:58.997Z",
          "content": "<p><a href=\"/umbertogriffo\">@umbertogriffo</a> Thanks much!</p>",
          "rawMarkdown": "@umbertogriffo Thanks much!"
        },
        {
          "id": 496818,
          "postDate": "2019-03-22T16:09:01.787Z",
          "content": "<p>Is the Densenet injection similar to resnet and resnext? </p>\n\n<p><code>\nx = stack1(x, 64, 3, stride1=1, name='conv2')\nx = stack1(x, 128, 4, name='conv3')\nx = stack1(x, 256, 6, name='conv4')\nx = stack1(x, 512, 3, name='conv5'\n</code></p>",
          "rawMarkdown": "Is the Densenet injection similar to resnet and resnext? \n\n```    \nx = stack1(x, 64, 3, stride1=1, name='conv2')\nx = stack1(x, 128, 4, name='conv3')\nx = stack1(x, 256, 6, name='conv4')\nx = stack1(x, 512, 3, name='conv5'\n```"
        }
      ]
    },
    {
      "id": 485671,
      "postDate": "2019-03-07T19:01:43.130Z",
      "content": "<p>A God among mortals. Ty</p>",
      "rawMarkdown": "A God among mortals. Ty"
    },
    {
      "id": 484615,
      "postDate": "2019-03-06T08:32:25.560Z",
      "content": "<p><a href=\"/franchini\">@franchini</a> many thanks! If you click reconnect colab resume the program execution.</p>",
      "rawMarkdown": "@franchini many thanks! If you click reconnect colab resume the program execution."
    },
    {
      "id": 484607,
      "postDate": "2019-03-06T08:28:15.090Z",
      "content": "<p><a href=\"/franky12\">@franky12</a> You are welcome! I rewritten ResNet50/ResNeXt50 models because the original implementation threw several exceptions when the code was compiled for the TPU.</p>",
      "rawMarkdown": "@franky12 You are welcome! I rewritten ResNet50/ResNeXt50 models because the original implementation threw several exceptions when the code was compiled for the TPU.",
      "replies": [
        {
          "id": 484623,
          "postDate": "2019-03-06T08:41:29.923Z",
          "content": "<p>Okay, thanks for the clarification!</p>",
          "rawMarkdown": "Okay, thanks for the clarification!"
        },
        {
          "id": 484838,
          "postDate": "2019-03-06T15:02:31.073Z",
          "content": "<p>Have u made any modification to the layers of ResNet50/ResNeXt50 model?</p>",
          "rawMarkdown": "Have u made any modification to the layers of ResNet50/ResNeXt50 model?"
        }
      ]
    },
    {
      "id": 484572,
      "postDate": "2019-03-06T07:32:57.950Z",
      "content": "<p>Thanks for sharing!\nWhy did you wrote your own ResNet50/ResNeXt50 models?\nThey are already available in <code>keras.applications</code></p>",
      "rawMarkdown": "Thanks for sharing!\nWhy did you wrote your own ResNet50/ResNeXt50 models?\nThey are already available in `keras.applications`"
    },
    {
      "id": 485300,
      "postDate": "2019-03-07T07:46:26.573Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 484767,
      "postDate": "2019-03-06T12:57:48.133Z",
      "content": "<p>Thanks for sharing!!</p>",
      "rawMarkdown": "Thanks for sharing!!",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 483472,
      "author_name": "Franchini",
      "author_url": "",
      "post_date": "2019-03-04T16:49:39.207000",
      "content": "<p>Thanks for sharing Umberto!\nHow do you keep Colab from disconnecting due to user inactivity?</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 487865,
      "author_name": "Appy Patel",
      "author_url": "",
      "post_date": "2019-03-11T15:11:36.347000",
      "content": "<p>Using which cnn model have u got 0.9757 score?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 488926,
          "author_name": "Umberto",
          "author_url": "",
          "post_date": "2019-03-13T08:14:11.313000",
          "content": "<p><a href=\"/aaryapatel\">@aaryapatel</a> I used different models and approaches in order to got 0.9757 score. I used an ensemble of  ResNet50, ResNext50, Densenet121, DenseNet169 and LGBM (VGG16 as feature extractor).</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 490253,
          "author_name": "Aarya Patel",
          "author_url": "",
          "post_date": "2019-03-14T12:37:58.997000",
          "content": "<p><a href=\"/umbertogriffo\">@umbertogriffo</a> Thanks much!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 496818,
          "author_name": "William Green",
          "author_url": "",
          "post_date": "2019-03-22T16:09:01.787000",
          "content": "<p>Is the Densenet injection similar to resnet and resnext? </p>\n\n<p><code>\nx = stack1(x, 64, 3, stride1=1, name='conv2')\nx = stack1(x, 128, 4, name='conv3')\nx = stack1(x, 256, 6, name='conv4')\nx = stack1(x, 512, 3, name='conv5'\n</code></p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 485671,
      "author_name": "Spiros Vondopoulos",
      "author_url": "",
      "post_date": "2019-03-07T19:01:43.130000",
      "content": "<p>A God among mortals. Ty</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 484615,
      "author_name": "Umberto",
      "author_url": "",
      "post_date": "2019-03-06T08:32:25.560000",
      "content": "<p><a href=\"/franchini\">@franchini</a> many thanks! If you click reconnect colab resume the program execution.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 484607,
      "author_name": "Umberto",
      "author_url": "",
      "post_date": "2019-03-06T08:28:15.090000",
      "content": "<p><a href=\"/franky12\">@franky12</a> You are welcome! I rewritten ResNet50/ResNeXt50 models because the original implementation threw several exceptions when the code was compiled for the TPU.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 484623,
          "author_name": "Arvid Teichtmann",
          "author_url": "",
          "post_date": "2019-03-06T08:41:29.923000",
          "content": "<p>Okay, thanks for the clarification!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 484838,
          "author_name": "Appy Patel",
          "author_url": "",
          "post_date": "2019-03-06T15:02:31.073000",
          "content": "<p>Have u made any modification to the layers of ResNet50/ResNeXt50 model?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 484572,
      "author_name": "Arvid Teichtmann",
      "author_url": "",
      "post_date": "2019-03-06T07:32:57.950000",
      "content": "<p>Thanks for sharing!\nWhy did you wrote your own ResNet50/ResNeXt50 models?\nThey are already available in <code>keras.applications</code></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 485300,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-03-07T07:46:26.573000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 484767,
      "author_name": "Rocky Xu",
      "author_url": "",
      "post_date": "2019-03-06T12:57:48.133000",
      "content": "<p>Thanks for sharing!!</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "483355": "Colab jupyter notebook on how to train a ResNet50/ResNeXt50 implemented on Keras with TPU on Google colab for free. \n[Check it out!]( https://tinyurl.com/y54y2xs9)",
    "483472": "Thanks for sharing Umberto!\nHow do you keep Colab from disconnecting due to user inactivity?",
    "487865": "Using which cnn model have u got 0.9757 score?",
    "485671": "A God among mortals. Ty",
    "484615": "@franchini many thanks! If you click reconnect colab resume the program execution.",
    "484607": "@franky12 You are welcome! I rewritten ResNet50/ResNeXt50 models because the original implementation threw several exceptions when the code was compiled for the TPU.",
    "484572": "Thanks for sharing!\nWhy did you wrote your own ResNet50/ResNeXt50 models?\nThey are already available in `keras.applications`",
    "485300": "",
    "484767": "Thanks for sharing!!"
  }
}