{
  "id": 314997,
  "title": "training time",
  "url": "/competitions/happy-whale-and-dolphin/discussion/314997",
  "author_name": "",
  "post_date": "2022-03-25T17:46:44.013890300Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi all👋<br>\niam new in this competition sorry if my question seems simple😊<br>\nI am using the fastai to train the efficientnet_b0 model, and I noticed that it takes about an hour for each epoch, while there are kernels whose training does not exceed ten minutes per epoch using same model with arcface loss and the same way to split data.I don't know if the bug is in the fast ai library or not I've tried to freeze the model layers as well as reduce the precision in the calculations  to 16 bit, but all this did not work??</p>\n<p>Note// when I use the CPU the training period is 2 hours per epoch.</p>",
  "messages": [
    {
      "id": "1734919",
      "postDate": "03/25/2022 17:46:44",
      "content": "<p>Hi all👋<br>\niam new in this competition sorry if my question seems simple😊<br>\nI am using the fastai to train the efficientnet_b0 model, and I noticed that it takes about an hour for each epoch, while there are kernels whose training does not exceed ten minutes per epoch using same model with arcface loss and the same way to split data.I don't know if the bug is in the fast ai library or not I've tried to freeze the model layers as well as reduce the precision in the calculations  to 16 bit, but all this did not work??</p>\n<p>Note// when I use the CPU the training period is 2 hours per epoch.</p>",
      "rawMarkdown": "Hi all👋\niam new in this competition sorry if my question seems simple😊\nI am using the fastai to train the efficientnet_b0 model, and I noticed that it takes about an hour for each epoch, while there are kernels whose training does not exceed ten minutes per epoch using same model with arcface loss and the same way to split data.I don't know if the bug is in the fast ai library or not I've tried to freeze the model layers as well as reduce the precision in the calculations  to 16 bit, but all this did not work??\n\nNote// when I use the CPU the training period is 2 hours per epoch.",
      "votes": null
    },
    {
      "id": "1735133",
      "postDate": "03/25/2022 22:18:39",
      "content": "<p>Try to use the TPU resource provided by kaggle which is much faster, 8 to 10 min an epoch</p>",
      "rawMarkdown": "Try to use the TPU resource provided by kaggle which is much faster, 8 to 10 min an epoch",
      "votes": null
    },
    {
      "id": "1735609",
      "postDate": "03/26/2022 12:43:59",
      "content": "<p><a href=\"https://www.kaggle.com/riadalmadani\" target=\"_blank\">@riadalmadani</a> I believe that you should enable the GPU accelerator. You can do it in the tab on the right when working with the notebook kernel. Also, the TPU accelerator is much faster than the GPU. Here, check out <a href=\"https://www.kaggle.com/code/ks2019/happywhale-arcface-baseline-tpu\" target=\"_blank\">this great notebook with TPU usage</a> in this competition</p>",
      "rawMarkdown": "riadalmadani I believe that you should enable the GPU accelerator. You can do it in the tab on the right when working with the notebook kernel. Also, the TPU accelerator is much faster than the GPU. Here, check out [this great notebook with TPU usage](https://www.kaggle.com/code/ks2019/happywhale-arcface-baseline-tpu) in this competition",
      "votes": null
    },
    {
      "id": "1735619",
      "postDate": "03/26/2022 12:51:08",
      "content": "<p>Thank you i think that i figured out the problem thanks again 😊</p>",
      "rawMarkdown": "Thank you i think that i figured out the problem thanks again 😊",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1735133,
      "author_name": "runjiali",
      "author_url": "",
      "post_date": "03/25/2022 22:18:39",
      "content": "<p>Try to use the TPU resource provided by kaggle which is much faster, 8 to 10 min an epoch</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1735609,
      "author_name": "ivanovakm",
      "author_url": "",
      "post_date": "03/26/2022 12:43:59",
      "content": "<p><a href=\"https://www.kaggle.com/riadalmadani\" target=\"_blank\">@riadalmadani</a> I believe that you should enable the GPU accelerator. You can do it in the tab on the right when working with the notebook kernel. Also, the TPU accelerator is much faster than the GPU. Here, check out <a href=\"https://www.kaggle.com/code/ks2019/happywhale-arcface-baseline-tpu\" target=\"_blank\">this great notebook with TPU usage</a> in this competition</p>",
      "votes": null,
      "replies": [
        {
          "id": 1735619,
          "author_name": "riadalmadani",
          "author_url": "",
          "post_date": "03/26/2022 12:51:08",
          "content": "<p>Thank you i think that i figured out the problem thanks again 😊</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1734919": "Hi all👋\niam new in this competition sorry if my question seems simple😊\nI am using the fastai to train the efficientnet_b0 model, and I noticed that it takes about an hour for each epoch, while there are kernels whose training does not exceed ten minutes per epoch using same model with arcface loss and the same way to split data.I don't know if the bug is in the fast ai library or not I've tried to freeze the model layers as well as reduce the precision in the calculations  to 16 bit, but all this did not work??\n\nNote// when I use the CPU the training period is 2 hours per epoch.",
    "1735133": "Try to use the TPU resource provided by kaggle which is much faster, 8 to 10 min an epoch",
    "1735609": "riadalmadani I believe that you should enable the GPU accelerator. You can do it in the tab on the right when working with the notebook kernel. Also, the TPU accelerator is much faster than the GPU. Here, check out [this great notebook with TPU usage](https://www.kaggle.com/code/ks2019/happywhale-arcface-baseline-tpu) in this competition",
    "1735619": "Thank you i think that i figured out the problem thanks again 😊"
  },
  "source": "meta"
}