{
  "id": 248683,
  "title": "Real time data augmentation on TPU",
  "url": "/competitions/siim-covid19-detection/discussion/248683",
  "author_name": "",
  "post_date": "2021-06-24T13:43:21.701555800Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>It seems that TPU only accepts tf.Dataset and according to the answer to this thread: <a href=\"https://stackoverflow.com/questions/60744247/augmentation-of-a-tf-data-dataset\" target=\"_blank\">https://stackoverflow.com/questions/60744247/augmentation-of-a-tf-data-dataset</a> augmentation is done only once. Furthermore effnetB7 does not fit into the memory of the single GPU available on Kaggle. Does somebody have a solution to do real-time data augmentation on TPU?</p>",
  "messages": [
    {
      "id": "1363936",
      "postDate": "06/24/2021 13:43:21",
      "content": "<p>It seems that TPU only accepts tf.Dataset and according to the answer to this thread: <a href=\"https://stackoverflow.com/questions/60744247/augmentation-of-a-tf-data-dataset\" target=\"_blank\">https://stackoverflow.com/questions/60744247/augmentation-of-a-tf-data-dataset</a> augmentation is done only once. Furthermore effnetB7 does not fit into the memory of the single GPU available on Kaggle. Does somebody have a solution to do real-time data augmentation on TPU?</p>",
      "rawMarkdown": "It seems that TPU only accepts tf.Dataset and according to the answer to this thread: https://stackoverflow.com/questions/60744247/augmentation-of-a-tf-data-dataset augmentation is done only once. Furthermore effnetB7 does not fit into the memory of the single GPU available on Kaggle. Does somebody have a solution to do real-time data augmentation on TPU?",
      "votes": null
    },
    {
      "id": "1364659",
      "postDate": "06/25/2021 06:00:37",
      "content": "<p>I think these notebooks by Chris do what you are looking for:</p>\n<ol>\n<li><a href=\"https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96\" target=\"_blank\">Rotation Augmentation GPU/TPU</a></li>\n<li><a href=\"https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu\" target=\"_blank\">CutMix and MixUp on GPU/TPU</a></li>\n</ol>",
      "rawMarkdown": "I think these notebooks by Chris do what you are looking for:\n\n1. [Rotation Augmentation GPU/TPU](https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96)\n2. [CutMix and MixUp on GPU/TPU](https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu)",
      "votes": null
    },
    {
      "id": "1364893",
      "postDate": "06/25/2021 09:37:25",
      "content": "<p>Thanks, it remains unclear for me if these notebooks do online augmentation or just augmentation at the start of the training.</p>",
      "rawMarkdown": "Thanks, it remains unclear for me if these notebooks do online augmentation or just augmentation at the start of the training.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1364659,
      "author_name": "moeinshariatnia",
      "author_url": "",
      "post_date": "06/25/2021 06:00:37",
      "content": "<p>I think these notebooks by Chris do what you are looking for:</p>\n<ol>\n<li><a href=\"https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96\" target=\"_blank\">Rotation Augmentation GPU/TPU</a></li>\n<li><a href=\"https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu\" target=\"_blank\">CutMix and MixUp on GPU/TPU</a></li>\n</ol>",
      "votes": null,
      "replies": [
        {
          "id": 1364893,
          "author_name": "josephamigo",
          "author_url": "",
          "post_date": "06/25/2021 09:37:25",
          "content": "<p>Thanks, it remains unclear for me if these notebooks do online augmentation or just augmentation at the start of the training.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1363936": "It seems that TPU only accepts tf.Dataset and according to the answer to this thread: https://stackoverflow.com/questions/60744247/augmentation-of-a-tf-data-dataset augmentation is done only once. Furthermore effnetB7 does not fit into the memory of the single GPU available on Kaggle. Does somebody have a solution to do real-time data augmentation on TPU?",
    "1364659": "I think these notebooks by Chris do what you are looking for:\n\n1. [Rotation Augmentation GPU/TPU](https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96)\n2. [CutMix and MixUp on GPU/TPU](https://www.kaggle.com/cdeotte/cutmix-and-mixup-on-gpu-tpu)",
    "1364893": "Thanks, it remains unclear for me if these notebooks do online augmentation or just augmentation at the start of the training."
  },
  "source": "meta"
}