{
  "id": 111839,
  "title": "Efficient Net B4-B7 ",
  "url": "/competitions/understanding_cloud_organization/discussion/111839",
  "author_name": "",
  "post_date": "2019-10-09T11:24:27.337469300Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi, </p>\n\n<p>I was trying to run EfficientnetBx (X = 4 to 7) using the below kernel </p>\n\n<p><a href=\"https://www.kaggle.com/datachampion/keras-efficientnetb3-for-classifying-cloud/edit\">https://www.kaggle.com/datachampion/keras-efficientnetb3-for-classifying-cloud/edit</a></p>\n\n<p>but I always encounter the OOM error for B4-B7. Would the reduced image size help? As of now its 300.</p>",
  "messages": [
    {
      "id": "644805",
      "postDate": "10/09/2019 11:24:27",
      "content": "<p>Hi, </p>\n\n<p>I was trying to run EfficientnetBx (X = 4 to 7) using the below kernel </p>\n\n<p><a href=\"https://www.kaggle.com/datachampion/keras-efficientnetb3-for-classifying-cloud/edit\">https://www.kaggle.com/datachampion/keras-efficientnetb3-for-classifying-cloud/edit</a></p>\n\n<p>but I always encounter the OOM error for B4-B7. Would the reduced image size help? As of now its 300.</p>",
      "rawMarkdown": "Hi, \n\nI was trying to run EfficientnetBx (X = 4 to 7) using the below kernel \n\nhttps://www.kaggle.com/datachampion/keras-efficientnetb3-for-classifying-cloud/edit\n\nbut I always encounter the OOM error for B4-B7. Would the reduced image size help? As of now its 300.",
      "votes": null
    },
    {
      "id": "644859",
      "postDate": "10/09/2019 12:42:27",
      "content": "<p>Try smaller batch sizes.</p>",
      "rawMarkdown": "Try smaller batch sizes.",
      "votes": null
    },
    {
      "id": "644919",
      "postDate": "10/09/2019 14:22:05",
      "content": "<p>you can try this customized callback, may help:\n<a href=\"https://www.kaggle.com/c/understanding_cloud_organization/discussion/111514#latest-642878\">https://www.kaggle.com/c/understanding_cloud_organization/discussion/111514#latest-642878</a></p>",
      "rawMarkdown": "you can try this customized callback, may help:\nhttps://www.kaggle.com/c/understanding_cloud_organization/discussion/111514#latest-642878",
      "votes": null
    },
    {
      "id": "645143",
      "postDate": "10/09/2019 20:53:08",
      "content": "<p>It might also be worthwhile to check the following <a href=\"https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/111292\">thread.</a> You can achieve 20% memory reduction, changing the definition of the Swish activation function</p>",
      "rawMarkdown": "It might also be worthwhile to check the following [thread.](https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/111292) You can achieve 20% memory reduction, changing the definition of the Swish activation function",
      "votes": null
    },
    {
      "id": "645420",
      "postDate": "10/10/2019 05:39:52",
      "content": "<p><a href=\"/gogo827jz\">@gogo827jz</a> , Thanks for the suggestion.\nIt helped to run the algo, but the accuracy dropped marginally :(</p>",
      "rawMarkdown": "gogo827jz , Thanks for the suggestion.\nIt helped to run the algo, but the accuracy dropped marginally :(",
      "votes": null
    },
    {
      "id": "646019",
      "postDate": "10/10/2019 18:34:01",
      "content": "<p>Tried with batch size 16 on B4 and image height and width = 256.  The result is stuck on LB to 0.655 \nBelow is the link of the Kernel\n<a href=\"https://www.kaggle.com/datachampion/keras-efficientnetb4\">https://www.kaggle.com/datachampion/keras-efficientnetb4</a></p>",
      "rawMarkdown": "Tried with batch size 16 on B4 and image height and width = 256.  The result is stuck on LB to 0.655 \nBelow is the link of the Kernel\nhttps://www.kaggle.com/datachampion/keras-efficientnetb4",
      "votes": null
    },
    {
      "id": "646176",
      "postDate": "10/10/2019 23:40:59",
      "content": "<p>Meanwhile, you may need gradient accumulation to overcome the small-batch drawback.</p>",
      "rawMarkdown": "Meanwhile, you may need gradient accumulation to overcome the small-batch drawback.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 644859,
      "author_name": "gogo827jz",
      "author_url": "",
      "post_date": "10/09/2019 12:42:27",
      "content": "<p>Try smaller batch sizes.</p>",
      "votes": null,
      "replies": [
        {
          "id": 645420,
          "author_name": "datachampion",
          "author_url": "",
          "post_date": "10/10/2019 05:39:52",
          "content": "<p><a href=\"/gogo827jz\">@gogo827jz</a> , Thanks for the suggestion.\nIt helped to run the algo, but the accuracy dropped marginally :(</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 646176,
          "author_name": "gogo827jz",
          "author_url": "",
          "post_date": "10/10/2019 23:40:59",
          "content": "<p>Meanwhile, you may need gradient accumulation to overcome the small-batch drawback.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 644919,
      "author_name": "anuragtr",
      "author_url": "",
      "post_date": "10/09/2019 14:22:05",
      "content": "<p>you can try this customized callback, may help:\n<a href=\"https://www.kaggle.com/c/understanding_cloud_organization/discussion/111514#latest-642878\">https://www.kaggle.com/c/understanding_cloud_organization/discussion/111514#latest-642878</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 645143,
      "author_name": "maxjeblick",
      "author_url": "",
      "post_date": "10/09/2019 20:53:08",
      "content": "<p>It might also be worthwhile to check the following <a href=\"https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/111292\">thread.</a> You can achieve 20% memory reduction, changing the definition of the Swish activation function</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 646019,
      "author_name": "datachampion",
      "author_url": "",
      "post_date": "10/10/2019 18:34:01",
      "content": "<p>Tried with batch size 16 on B4 and image height and width = 256.  The result is stuck on LB to 0.655 \nBelow is the link of the Kernel\n<a href=\"https://www.kaggle.com/datachampion/keras-efficientnetb4\">https://www.kaggle.com/datachampion/keras-efficientnetb4</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "644805": "Hi, \n\nI was trying to run EfficientnetBx (X = 4 to 7) using the below kernel \n\nhttps://www.kaggle.com/datachampion/keras-efficientnetb3-for-classifying-cloud/edit\n\nbut I always encounter the OOM error for B4-B7. Would the reduced image size help? As of now its 300.",
    "644859": "Try smaller batch sizes.",
    "644919": "you can try this customized callback, may help:\nhttps://www.kaggle.com/c/understanding_cloud_organization/discussion/111514#latest-642878",
    "645143": "It might also be worthwhile to check the following [thread.](https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/111292) You can achieve 20% memory reduction, changing the definition of the Swish activation function",
    "645420": "gogo827jz , Thanks for the suggestion.\nIt helped to run the algo, but the accuracy dropped marginally :(",
    "646019": "Tried with batch size 16 on B4 and image height and width = 256.  The result is stuck on LB to 0.655 \nBelow is the link of the Kernel\nhttps://www.kaggle.com/datachampion/keras-efficientnetb4",
    "646176": "Meanwhile, you may need gradient accumulation to overcome the small-batch drawback."
  },
  "source": "meta"
}