{
  "id": 111880,
  "title": "what opt_level would you choose for Apex?",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/111880",
  "author_name": "Salaryman",
  "post_date": "2019-10-09T16:07:50.072000",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi I am new to Apex for training, may I know if I have a RTX GPU, should I go for 01(Mixed Precision) or 02(Almost FP16 Mixed Precision)?</p>\n\n<p>From the <a href=\"https://nvidia.github.io/apex/amp.html\">doc</a> it seems 01 provides more stability but 02 can train faster. </p>\n\n<p>Anyone can share some experience about the speed and accuracy trade off??\nThanks so much</p>",
  "messages": [
    {
      "id": 644989,
      "postDate": "2019-10-09T16:07:50.073Z",
      "content": "<p>Hi I am new to Apex for training, may I know if I have a RTX GPU, should I go for 01(Mixed Precision) or 02(Almost FP16 Mixed Precision)?</p>\n\n<p>From the <a href=\"https://nvidia.github.io/apex/amp.html\">doc</a> it seems 01 provides more stability but 02 can train faster. </p>\n\n<p>Anyone can share some experience about the speed and accuracy trade off??\nThanks so much</p>",
      "rawMarkdown": "Hi I am new to Apex for training, may I know if I have a RTX GPU, should I go for 01(Mixed Precision) or 02(Almost FP16 Mixed Precision)?\n\nFrom the [doc](https://nvidia.github.io/apex/amp.html) it seems 01 provides more stability but 02 can train faster. \n\nAnyone can share some experience about the speed and accuracy trade off??\nThanks so much\n",
      "votes": 1
    },
    {
      "id": 645128,
      "postDate": "2019-10-09T20:37:45.487Z",
      "content": "<p>It depends on your model and other things like your preprocessing, data loading etc. I made quite different experience in different competition, so I advise to try both and compare. </p>",
      "rawMarkdown": "It depends on your model and other things like your preprocessing, data loading etc. I made quite different experience in different competition, so I advise to try both and compare. ",
      "votes": 2,
      "replies": [
        {
          "id": 645865,
          "postDate": "2019-10-10T14:59:29.407Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 647437,
          "postDate": "2019-10-12T16:08:27.120Z",
          "content": "<p>thanks so much! Im trying O2 now</p>",
          "rawMarkdown": "thanks so much! Im trying O2 now"
        }
      ]
    },
    {
      "id": 646381,
      "postDate": "2019-10-11T07:38:19.420Z",
      "content": "<p>i go with O2 and the accuracy tradeoff is really small</p>",
      "rawMarkdown": "i go with O2 and the accuracy tradeoff is really small",
      "replies": [
        {
          "id": 647439,
          "postDate": "2019-10-12T16:10:34.627Z",
          "content": "<p>Thanks! I don't have to worry about training with O2 after seeing your score🙌 </p>",
          "rawMarkdown": "Thanks! I don't have to worry about training with O2 after seeing your score🙌 "
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 645128,
      "author_name": "Dieter",
      "author_url": "",
      "post_date": "2019-10-09T20:37:45.487000",
      "content": "<p>It depends on your model and other things like your preprocessing, data loading etc. I made quite different experience in different competition, so I advise to try both and compare. </p>",
      "votes": 2,
      "replies": [
        {
          "id": 645865,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-10-10T14:59:29.407000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 647437,
          "author_name": "Salaryman",
          "author_url": "",
          "post_date": "2019-10-12T16:08:27.120000",
          "content": "<p>thanks so much! Im trying O2 now</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 646381,
      "author_name": "DatNT",
      "author_url": "",
      "post_date": "2019-10-11T07:38:19.420000",
      "content": "<p>i go with O2 and the accuracy tradeoff is really small</p>",
      "votes": 0,
      "replies": [
        {
          "id": 647439,
          "author_name": "Salaryman",
          "author_url": "",
          "post_date": "2019-10-12T16:10:34.627000",
          "content": "<p>Thanks! I don't have to worry about training with O2 after seeing your score🙌 </p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "644989": "Hi I am new to Apex for training, may I know if I have a RTX GPU, should I go for 01(Mixed Precision) or 02(Almost FP16 Mixed Precision)?\n\nFrom the [doc](https://nvidia.github.io/apex/amp.html) it seems 01 provides more stability but 02 can train faster. \n\nAnyone can share some experience about the speed and accuracy trade off??\nThanks so much\n",
    "645128": "It depends on your model and other things like your preprocessing, data loading etc. I made quite different experience in different competition, so I advise to try both and compare. ",
    "646381": "i go with O2 and the accuracy tradeoff is really small"
  }
}