{
  "id": 133696,
  "title": "Video reader",
  "url": "/competitions/deepfake-detection-challenge/discussion/133696",
  "author_name": "",
  "post_date": "2020-03-03T23:28:20.805205200Z",
  "votes": 8,
  "comment_count": 11,
  "views": 0,
  "content": "<p>I try few video reader. Current my best reader is decord-gpu\n<a href=\"https://www.kaggle.com/leighplt/decord-videoreader\">https://www.kaggle.com/leighplt/decord-videoreader</a>\ncpu version read by 20 min this 400 files.\nHow found more faster framework?</p>",
  "messages": [
    {
      "id": "762898",
      "postDate": "03/03/2020 23:28:20",
      "content": "<p>I try few video reader. Current my best reader is decord-gpu\n<a href=\"https://www.kaggle.com/leighplt/decord-videoreader\">https://www.kaggle.com/leighplt/decord-videoreader</a>\ncpu version read by 20 min this 400 files.\nHow found more faster framework?</p>",
      "rawMarkdown": "I try few video reader. Current my best reader is decord-gpu\nhttps://www.kaggle.com/leighplt/decord-videoreader\ncpu version read by 20 min this 400 files.\nHow found more faster framework?",
      "votes": null
    },
    {
      "id": "762906",
      "postDate": "03/03/2020 23:40:33",
      "content": "<p>this is interesting... was just starting to look into this. can you make the discord dataset public so i can try the kernel?</p>",
      "rawMarkdown": "this is interesting... was just starting to look into this. can you make the discord dataset public so i can try the kernel?",
      "votes": null
    },
    {
      "id": "762911",
      "postDate": "03/03/2020 23:47:43",
      "content": "<p>shared</p>",
      "rawMarkdown": "shared",
      "votes": null
    },
    {
      "id": "762995",
      "postDate": "03/04/2020 02:13:28",
      "content": "<p>Thank you, it does seem to be much faster then anything I tried so far and very neat. If using keras/tf remember that tf likes to grab all gpu memory, so you need to tell it to softplace.</p>",
      "rawMarkdown": "Thank you, it does seem to be much faster then anything I tried so far and very neat. If using keras/tf remember that tf likes to grab all gpu memory, so you need to tell it to softplace.",
      "votes": null
    },
    {
      "id": "763201",
      "postDate": "03/04/2020 08:27:07",
      "content": "<p>Decord used gpu memory not accurate with cuda API and we can see memory leak.</p>",
      "rawMarkdown": "Decord used gpu memory not accurate with cuda API and we can see memory leak.",
      "votes": null
    },
    {
      "id": "763212",
      "postDate": "03/04/2020 08:50:22",
      "content": "<p>yes, i also thought it leaked a bit, but for the amounts we use here its not a big deal. I think it was like few mb per 400 files. I wasn't sure if it was a leak or it just consumes the most it needs (like for higher resolution movies).</p>",
      "rawMarkdown": "yes, i also thought it leaked a bit, but for the amounts we use here its not a big deal. I think it was like few mb per 400 files. I wasn't sure if it was a leak or it just consumes the most it needs (like for higher resolution movies).",
      "votes": null
    },
    {
      "id": "763602",
      "postDate": "03/04/2020 16:49:29",
      "content": "<p>Be carefull with gpu - memory leak. Use CPU its stable and faster than cv2</p>",
      "rawMarkdown": "Be carefull with gpu - memory leak. Use CPU its stable and faster than cv2",
      "votes": null
    },
    {
      "id": "763895",
      "postDate": "03/05/2020 00:28:29",
      "content": "<p>Thank you!</p>",
      "rawMarkdown": "Thank you!",
      "votes": null
    },
    {
      "id": "763963",
      "postDate": "03/05/2020 03:09:59",
      "content": "<p>For GPU memory leak consider using <a href=\"https://github.com/stas00/ipyexperiments\">https://github.com/stas00/ipyexperiments</a>. You can create contexts and run your code inside it, later code is complete all memory will be claimed back with garbage collection. Haven't tested it with Decord yet.</p>",
      "rawMarkdown": "For GPU memory leak consider using https://github.com/stas00/ipyexperiments. You can create contexts and run your code inside it, later code is complete all memory will be claimed back with garbage collection. Haven't tested it with Decord yet.",
      "votes": null
    },
    {
      "id": "771960",
      "postDate": "03/14/2020 21:13:17",
      "content": "<p>yeah, there's ram leak on gpu version.</p>",
      "rawMarkdown": "yeah, there's ram leak on gpu version.",
      "votes": null
    },
    {
      "id": "784372",
      "postDate": "03/24/2020 06:55:47",
      "content": "<p>How to use it on cpu, with the accelerator turned off? My GPU hours finished, and I cannot even install the package from your notebook (missing files with nvidia library)?</p>",
      "rawMarkdown": "How to use it on cpu, with the accelerator turned off? My GPU hours finished, and I cannot even install the package from your notebook (missing files with nvidia library)?",
      "votes": null
    },
    {
      "id": "784735",
      "postDate": "03/24/2020 13:17:56",
      "content": "<p><code>pip install decord</code>\npip vertion w/o gpu</p>",
      "rawMarkdown": "`pip install decord`\npip vertion w/o gpu",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 762906,
      "author_name": "moshel",
      "author_url": "",
      "post_date": "03/03/2020 23:40:33",
      "content": "<p>this is interesting... was just starting to look into this. can you make the discord dataset public so i can try the kernel?</p>",
      "votes": null,
      "replies": [
        {
          "id": 762911,
          "author_name": "leighplt",
          "author_url": "",
          "post_date": "03/03/2020 23:47:43",
          "content": "<p>shared</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 762995,
          "author_name": "moshel",
          "author_url": "",
          "post_date": "03/04/2020 02:13:28",
          "content": "<p>Thank you, it does seem to be much faster then anything I tried so far and very neat. If using keras/tf remember that tf likes to grab all gpu memory, so you need to tell it to softplace.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 763201,
      "author_name": "leighplt",
      "author_url": "",
      "post_date": "03/04/2020 08:27:07",
      "content": "<p>Decord used gpu memory not accurate with cuda API and we can see memory leak.</p>",
      "votes": null,
      "replies": [
        {
          "id": 763212,
          "author_name": "moshel",
          "author_url": "",
          "post_date": "03/04/2020 08:50:22",
          "content": "<p>yes, i also thought it leaked a bit, but for the amounts we use here its not a big deal. I think it was like few mb per 400 files. I wasn't sure if it was a leak or it just consumes the most it needs (like for higher resolution movies).</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 763602,
      "author_name": "leighplt",
      "author_url": "",
      "post_date": "03/04/2020 16:49:29",
      "content": "<p>Be carefull with gpu - memory leak. Use CPU its stable and faster than cv2</p>",
      "votes": null,
      "replies": [
        {
          "id": 763895,
          "author_name": "moshel",
          "author_url": "",
          "post_date": "03/05/2020 00:28:29",
          "content": "<p>Thank you!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 771960,
          "author_name": "yimacs",
          "author_url": "",
          "post_date": "03/14/2020 21:13:17",
          "content": "<p>yeah, there's ram leak on gpu version.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 763963,
      "author_name": "keremt",
      "author_url": "",
      "post_date": "03/05/2020 03:09:59",
      "content": "<p>For GPU memory leak consider using <a href=\"https://github.com/stas00/ipyexperiments\">https://github.com/stas00/ipyexperiments</a>. You can create contexts and run your code inside it, later code is complete all memory will be claimed back with garbage collection. Haven't tested it with Decord yet.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 784372,
      "author_name": "fatyild",
      "author_url": "",
      "post_date": "03/24/2020 06:55:47",
      "content": "<p>How to use it on cpu, with the accelerator turned off? My GPU hours finished, and I cannot even install the package from your notebook (missing files with nvidia library)?</p>",
      "votes": null,
      "replies": [
        {
          "id": 784735,
          "author_name": "leighplt",
          "author_url": "",
          "post_date": "03/24/2020 13:17:56",
          "content": "<p><code>pip install decord</code>\npip vertion w/o gpu</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "762898": "I try few video reader. Current my best reader is decord-gpu\nhttps://www.kaggle.com/leighplt/decord-videoreader\ncpu version read by 20 min this 400 files.\nHow found more faster framework?",
    "762906": "this is interesting... was just starting to look into this. can you make the discord dataset public so i can try the kernel?",
    "762911": "shared",
    "762995": "Thank you, it does seem to be much faster then anything I tried so far and very neat. If using keras/tf remember that tf likes to grab all gpu memory, so you need to tell it to softplace.",
    "763201": "Decord used gpu memory not accurate with cuda API and we can see memory leak.",
    "763212": "yes, i also thought it leaked a bit, but for the amounts we use here its not a big deal. I think it was like few mb per 400 files. I wasn't sure if it was a leak or it just consumes the most it needs (like for higher resolution movies).",
    "763602": "Be carefull with gpu - memory leak. Use CPU its stable and faster than cv2",
    "763895": "Thank you!",
    "763963": "For GPU memory leak consider using https://github.com/stas00/ipyexperiments. You can create contexts and run your code inside it, later code is complete all memory will be claimed back with garbage collection. Haven't tested it with Decord yet.",
    "771960": "yeah, there's ram leak on gpu version.",
    "784372": "How to use it on cpu, with the accelerator turned off? My GPU hours finished, and I cannot even install the package from your notebook (missing files with nvidia library)?",
    "784735": "`pip install decord`\npip vertion w/o gpu"
  },
  "source": "meta"
}