{
  "id": 319113,
  "title": "Why TPU baselines crashing on GPU?",
  "url": "/competitions/happy-whale-and-dolphin/discussion/319113",
  "author_name": "",
  "post_date": "2022-04-15T12:11:12.359206700Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I have tried lightweights models like efficient-net v1 B0, freeze backbone, set batch size 4. image size 512*512 (from happywhale-tfrecords-detic-box-512), but still crashes, I don't understand why.</p>\n<p>P.S. Seems like problem in <br>\ndataset.repeat() and   dataset.shuffle(2048) (to big shuffle buffer)</p>\n<p>P.S. I return dataset.repeat() because without it, model can't be fitted more than 1 epoch. Hope problem only with shuffle buffer to big</p>",
  "messages": [
    {
      "id": "1756329",
      "postDate": "04/15/2022 12:11:12",
      "content": "<p>I have tried lightweights models like efficient-net v1 B0, freeze backbone, set batch size 4. image size 512*512 (from happywhale-tfrecords-detic-box-512), but still crashes, I don't understand why.</p>\n<p>P.S. Seems like problem in <br>\ndataset.repeat() and   dataset.shuffle(2048) (to big shuffle buffer)</p>\n<p>P.S. I return dataset.repeat() because without it, model can't be fitted more than 1 epoch. Hope problem only with shuffle buffer to big</p>",
      "rawMarkdown": "I have tried lightweights models like efficient-net v1 B0, freeze backbone, set batch size 4. image size 512*512 (from happywhale-tfrecords-detic-box-512), but still crashes, I don't understand why.\n\nP.S. Seems like problem in \ndataset.repeat() and   dataset.shuffle(2048) (to big shuffle buffer)\n\nP.S. I return dataset.repeat() because without it, model can't be fitted more than 1 epoch. Hope problem only with shuffle buffer to big",
      "votes": null
    },
    {
      "id": "1757745",
      "postDate": "04/17/2022 01:27:31",
      "content": "<p>It will depend on your image size and batch size but, if I remember well, I could get the GPU to run an EfficientNetB4 or B5 by setting <code>dataset.shuffle(512)</code>. Hope this helps! </p>",
      "rawMarkdown": "It will depend on your image size and batch size but, if I remember well, I could get the GPU to run an EfficientNetB4 or B5 by setting `dataset.shuffle(512)`. Hope this helps!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1757745,
      "author_name": "frlemarchand",
      "author_url": "",
      "post_date": "04/17/2022 01:27:31",
      "content": "<p>It will depend on your image size and batch size but, if I remember well, I could get the GPU to run an EfficientNetB4 or B5 by setting <code>dataset.shuffle(512)</code>. Hope this helps! </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1756329": "I have tried lightweights models like efficient-net v1 B0, freeze backbone, set batch size 4. image size 512*512 (from happywhale-tfrecords-detic-box-512), but still crashes, I don't understand why.\n\nP.S. Seems like problem in \ndataset.repeat() and   dataset.shuffle(2048) (to big shuffle buffer)\n\nP.S. I return dataset.repeat() because without it, model can't be fitted more than 1 epoch. Hope problem only with shuffle buffer to big",
    "1757745": "It will depend on your image size and batch size but, if I remember well, I could get the GPU to run an EfficientNetB4 or B5 by setting `dataset.shuffle(512)`. Hope this helps!"
  },
  "source": "meta"
}