{
  "id": 359817,
  "title": "That time of the week",
  "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/359817",
  "author_name": "Samuel Cortinhas",
  "post_date": "2022-10-13T18:15:22.394000",
  "votes": 13,
  "comment_count": 3,
  "views": 0,
  "content": "<p><img src=\"https://i.postimg.cc/m2fGqQ1q/gpu-meme.jpg\"></p>\n<p>Someone please teach me how to use TPU with PyTorch…</p>",
  "messages": [
    {
      "id": 1985981,
      "postDate": "2022-10-13T18:15:22.393Z",
      "content": "<p><img src=\"https://i.postimg.cc/m2fGqQ1q/gpu-meme.jpg\"></p>\n<p>Someone please teach me how to use TPU with PyTorch…</p>",
      "rawMarkdown": "<img src='https://i.postimg.cc/m2fGqQ1q/gpu-meme.jpg' width=500>\n\nSomeone please teach me how to use TPU with PyTorch...",
      "votes": 13
    },
    {
      "id": 1996812,
      "postDate": "2022-10-20T12:42:37.837Z",
      "content": "<p>there is no way to fully use TPU power to train pytorch model.<br>\nnetwork IO would be the bottleneck.</p>",
      "rawMarkdown": "there is no way to fully use TPU power to train pytorch model.\nnetwork IO would be the bottleneck.",
      "votes": 1
    },
    {
      "id": 1992492,
      "postDate": "2022-10-17T17:41:04.353Z",
      "content": "<p>Certainly not the best time to rewrite your entire code, but check out pytorch lightning!</p>\n<p>For most of the time, it is as simple as installing some tpu related wheels and setting <code>accelerator=\"tpu\"</code> at training. </p>\n<p><a href=\"https://www.pytorchlightning.ai/\" target=\"_blank\">https://www.pytorchlightning.ai/</a></p>\n<p><a href=\"https://pytorch-lightning.readthedocs.io/en/stable/notebooks/lightning_examples/mnist-tpu-training.html\" target=\"_blank\">https://pytorch-lightning.readthedocs.io/en/stable/notebooks/lightning_examples/mnist-tpu-training.html</a></p>",
      "rawMarkdown": "Certainly not the best time to rewrite your entire code, but check out pytorch lightning!\n\nFor most of the time, it is as simple as installing some tpu related wheels and setting `accelerator=\"tpu\"` at training. \n\nhttps://www.pytorchlightning.ai/\n\nhttps://pytorch-lightning.readthedocs.io/en/stable/notebooks/lightning_examples/mnist-tpu-training.html",
      "votes": 1
    },
    {
      "id": 1986050,
      "postDate": "2022-10-13T19:06:27.247Z",
      "content": "<p>I think this is the state with most of us using our GPU quotas to the fullest. I think Kaggle can consider rolling over some hours of unused GPU quotas to the subsequent weeks if unused in the current week. I think this will help a lot of us.<br>\nFunny post <a href=\"https://www.kaggle.com/samuelcortinhas\" target=\"_blank\">@samuelcortinhas</a> </p>",
      "rawMarkdown": "I think this is the state with most of us using our GPU quotas to the fullest. I think Kaggle can consider rolling over some hours of unused GPU quotas to the subsequent weeks if unused in the current week. I think this will help a lot of us.\nFunny post @samuelcortinhas "
    }
  ],
  "comments": [
    {
      "id": 1996812,
      "author_name": "Qishen Ha",
      "author_url": "",
      "post_date": "2022-10-20T12:42:37.837000",
      "content": "<p>there is no way to fully use TPU power to train pytorch model.<br>\nnetwork IO would be the bottleneck.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1992492,
      "author_name": "SieunPark",
      "author_url": "",
      "post_date": "2022-10-17T17:41:04.353000",
      "content": "<p>Certainly not the best time to rewrite your entire code, but check out pytorch lightning!</p>\n<p>For most of the time, it is as simple as installing some tpu related wheels and setting <code>accelerator=\"tpu\"</code> at training. </p>\n<p><a href=\"https://www.pytorchlightning.ai/\" target=\"_blank\">https://www.pytorchlightning.ai/</a></p>\n<p><a href=\"https://pytorch-lightning.readthedocs.io/en/stable/notebooks/lightning_examples/mnist-tpu-training.html\" target=\"_blank\">https://pytorch-lightning.readthedocs.io/en/stable/notebooks/lightning_examples/mnist-tpu-training.html</a></p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1986050,
      "author_name": "Ravi Ramakrishnan",
      "author_url": "",
      "post_date": "2022-10-13T19:06:27.247000",
      "content": "<p>I think this is the state with most of us using our GPU quotas to the fullest. I think Kaggle can consider rolling over some hours of unused GPU quotas to the subsequent weeks if unused in the current week. I think this will help a lot of us.<br>\nFunny post <a href=\"https://www.kaggle.com/samuelcortinhas\" target=\"_blank\">@samuelcortinhas</a> </p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1985981": "<img src='https://i.postimg.cc/m2fGqQ1q/gpu-meme.jpg' width=500>\n\nSomeone please teach me how to use TPU with PyTorch...",
    "1996812": "there is no way to fully use TPU power to train pytorch model.\nnetwork IO would be the bottleneck.",
    "1992492": "Certainly not the best time to rewrite your entire code, but check out pytorch lightning!\n\nFor most of the time, it is as simple as installing some tpu related wheels and setting `accelerator=\"tpu\"` at training. \n\nhttps://www.pytorchlightning.ai/\n\nhttps://pytorch-lightning.readthedocs.io/en/stable/notebooks/lightning_examples/mnist-tpu-training.html",
    "1986050": "I think this is the state with most of us using our GPU quotas to the fullest. I think Kaggle can consider rolling over some hours of unused GPU quotas to the subsequent weeks if unused in the current week. I think this will help a lot of us.\nFunny post @samuelcortinhas "
  }
}