{
  "id": 49318,
  "title": "What kind of hardware are used?",
  "url": "/competitions/sp-society-camera-model-identification/discussion/49318",
  "author_name": "A0198918E_MaZhaoyang",
  "post_date": "2018-02-09T11:25:50.542000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Congratulations to all winners! Thanks Kaggle and IEEE Signal Processing Society for this competition!</p>\n\n<p>One question that I am curious and want to ask all winners is what hardware are being used. I am asking is because for me, to train each model, it takes a long time. Basically I need over 20 hours to train one model (either ResNet50, InceptionResnetV2, VGG16 or Xception). The input image size is cropped to 512x512 but I can only take a batch size of 4 (higher than this will cause a resource exhausted error). Even after 20 hours, the validation score never goes higher than 0.93.</p>\n\n<p>The hardware limitation is painful for me. </p>",
  "messages": [
    {
      "id": 280103,
      "postDate": "2018-02-09T11:25:50.543Z",
      "content": "<p>Congratulations to all winners! Thanks Kaggle and IEEE Signal Processing Society for this competition!</p>\n\n<p>One question that I am curious and want to ask all winners is what hardware are being used. I am asking is because for me, to train each model, it takes a long time. Basically I need over 20 hours to train one model (either ResNet50, InceptionResnetV2, VGG16 or Xception). The input image size is cropped to 512x512 but I can only take a batch size of 4 (higher than this will cause a resource exhausted error). Even after 20 hours, the validation score never goes higher than 0.93.</p>\n\n<p>The hardware limitation is painful for me. </p>",
      "rawMarkdown": "Congratulations to all winners! Thanks Kaggle and IEEE Signal Processing Society for this competition!\n\nOne question that I am curious and want to ask all winners is what hardware are being used. I am asking is because for me, to train each model, it takes a long time. Basically I need over 20 hours to train one model (either ResNet50, InceptionResnetV2, VGG16 or Xception). The input image size is cropped to 512x512 but I can only take a batch size of 4 (higher than this will cause a resource exhausted error). Even after 20 hours, the validation score never goes higher than 0.93.\n\nThe hardware limitation is painful for me. \n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "280103": "Congratulations to all winners! Thanks Kaggle and IEEE Signal Processing Society for this competition!\n\nOne question that I am curious and want to ask all winners is what hardware are being used. I am asking is because for me, to train each model, it takes a long time. Basically I need over 20 hours to train one model (either ResNet50, InceptionResnetV2, VGG16 or Xception). The input image size is cropped to 512x512 but I can only take a batch size of 4 (higher than this will cause a resource exhausted error). Even after 20 hours, the validation score never goes higher than 0.93.\n\nThe hardware limitation is painful for me. \n"
  }
}