{
  "id": 166253,
  "title": "EfficientNet- GPU or Tensors",
  "url": "/competitions/alaska2-image-steganalysis/discussion/166253",
  "author_name": "",
  "post_date": "2020-07-12T10:01:05.540002900Z",
  "votes": null,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi, trying out EffcientNet B7 on GPU and it is slow for training and as I try to train the model using different parameters. Running out of GPU hours</p>\n\n<p>Does it run faster on Tensors?</p>\n\n<p>Also, PyTorch runs on GPU only or can you use tensors with PyTorch</p>\n\n<p>Apologies for newbie question. Thanks for your guidance.</p>",
  "messages": [
    {
      "id": "925775",
      "postDate": "07/12/2020 10:01:05",
      "content": "<p>Hi, trying out EffcientNet B7 on GPU and it is slow for training and as I try to train the model using different parameters. Running out of GPU hours</p>\n\n<p>Does it run faster on Tensors?</p>\n\n<p>Also, PyTorch runs on GPU only or can you use tensors with PyTorch</p>\n\n<p>Apologies for newbie question. Thanks for your guidance.</p>",
      "rawMarkdown": "Hi, trying out EffcientNet B7 on GPU and it is slow for training and as I try to train the model using different parameters. Running out of GPU hours\n\nDoes it run faster on Tensors?\n\nAlso, PyTorch runs on GPU only or can you use tensors with PyTorch\n\nApologies for newbie question. Thanks for your guidance.",
      "votes": null
    },
    {
      "id": "925853",
      "postDate": "07/12/2020 10:51:10",
      "content": "<p>EfficentNet-B7 has huge number of parameters. if you can utilize TPUs of kaggle somehow, you will save some time I hope. </p>",
      "rawMarkdown": "EfficentNet-B7 has huge number of parameters. if you can utilize TPUs of kaggle somehow, you will save some time I hope.",
      "votes": null
    },
    {
      "id": "925900",
      "postDate": "07/12/2020 11:27:24",
      "content": "<p>Thanks <a href=\"/redwankarimsony\">@redwankarimsony</a> </p>\n\n<p>I am looking to try TPUs as well</p>",
      "rawMarkdown": "Thanks @redwankarimsony \n\nI am looking to try TPUs as well",
      "votes": null
    },
    {
      "id": "925972",
      "postDate": "07/12/2020 12:03:47",
      "content": "<p>Thanks <a href=\"/tarunpaparaju\">@tarunpaparaju</a> </p>\n\n<p>Found your notebook to use EfficientNet-B7 PyTorch in TPUs\nLink: <a href=\"https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch\">https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch</a></p>",
      "rawMarkdown": "Thanks @tarunpaparaju \n\nFound your notebook to use EfficientNet-B7 PyTorch in TPUs\nLink: [https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch](https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch)",
      "votes": null
    },
    {
      "id": "926396",
      "postDate": "07/12/2020 17:24:44",
      "content": "<p>I just want you to know that B7 on a TPU will take at least 12 hours to get to .90+ LB. </p>\n\n<p>You can also try the tf.keras implementation of efficientnet. <a href=\"https://www.kaggle.com/hooong/tpu-on-all-300k-images-without-crashing\">https://www.kaggle.com/hooong/tpu-on-all-300k-images-without-crashing</a></p>",
      "rawMarkdown": "I just want you to know that B7 on a TPU will take at least 12 hours to get to .90+ LB. \n\nYou can also try the tf.keras implementation of efficientnet. https://www.kaggle.com/hooong/tpu-on-all-300k-images-without-crashing",
      "votes": null
    },
    {
      "id": "931213",
      "postDate": "07/16/2020 04:15:32",
      "content": "<p>Thanks <a href=\"/hooong\">@hooong</a> for the kind advice and suggestion</p>",
      "rawMarkdown": "Thanks @hooong for the kind advice and suggestion",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 925853,
      "author_name": "redwankarimsony",
      "author_url": "",
      "post_date": "07/12/2020 10:51:10",
      "content": "<p>EfficentNet-B7 has huge number of parameters. if you can utilize TPUs of kaggle somehow, you will save some time I hope. </p>",
      "votes": null,
      "replies": [
        {
          "id": 925900,
          "author_name": "kmldas",
          "author_url": "",
          "post_date": "07/12/2020 11:27:24",
          "content": "<p>Thanks <a href=\"/redwankarimsony\">@redwankarimsony</a> </p>\n\n<p>I am looking to try TPUs as well</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 926396,
          "author_name": "hooong",
          "author_url": "",
          "post_date": "07/12/2020 17:24:44",
          "content": "<p>I just want you to know that B7 on a TPU will take at least 12 hours to get to .90+ LB. </p>\n\n<p>You can also try the tf.keras implementation of efficientnet. <a href=\"https://www.kaggle.com/hooong/tpu-on-all-300k-images-without-crashing\">https://www.kaggle.com/hooong/tpu-on-all-300k-images-without-crashing</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 931213,
          "author_name": "kmldas",
          "author_url": "",
          "post_date": "07/16/2020 04:15:32",
          "content": "<p>Thanks <a href=\"/hooong\">@hooong</a> for the kind advice and suggestion</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 925972,
      "author_name": "kmldas",
      "author_url": "",
      "post_date": "07/12/2020 12:03:47",
      "content": "<p>Thanks <a href=\"/tarunpaparaju\">@tarunpaparaju</a> </p>\n\n<p>Found your notebook to use EfficientNet-B7 PyTorch in TPUs\nLink: <a href=\"https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch\">https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "925775": "Hi, trying out EffcientNet B7 on GPU and it is slow for training and as I try to train the model using different parameters. Running out of GPU hours\n\nDoes it run faster on Tensors?\n\nAlso, PyTorch runs on GPU only or can you use tensors with PyTorch\n\nApologies for newbie question. Thanks for your guidance.",
    "925853": "EfficentNet-B7 has huge number of parameters. if you can utilize TPUs of kaggle somehow, you will save some time I hope.",
    "925900": "Thanks @redwankarimsony \n\nI am looking to try TPUs as well",
    "925972": "Thanks @tarunpaparaju \n\nFound your notebook to use EfficientNet-B7 PyTorch in TPUs\nLink: [https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch](https://www.kaggle.com/tarunpaparaju/alaska2-steganalysis-efficientnet-b3-pytorch)",
    "926396": "I just want you to know that B7 on a TPU will take at least 12 hours to get to .90+ LB. \n\nYou can also try the tf.keras implementation of efficientnet. https://www.kaggle.com/hooong/tpu-on-all-300k-images-without-crashing",
    "931213": "Thanks @hooong for the kind advice and suggestion"
  },
  "source": "meta"
}