{
  "id": 167065,
  "title": "What is best method to experiment",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/167065",
  "author_name": "",
  "post_date": "2020-07-15T04:51:41.627910800Z",
  "votes": 3,
  "comment_count": 5,
  "views": 0,
  "content": "<p>What is best method to experiment?\n( If I have 5 gpu, \n5 gpu using one experiment\nvs\neach gpu(1) using each experiment)</p>",
  "messages": [
    {
      "id": "929928",
      "postDate": "07/15/2020 04:51:41",
      "content": "<p>What is best method to experiment?\n( If I have 5 gpu, \n5 gpu using one experiment\nvs\neach gpu(1) using each experiment)</p>",
      "rawMarkdown": "What is best method to experiment?\n( If I have 5 gpu, \n5 gpu using one experiment\nvs\neach gpu(1) using each experiment)",
      "votes": null
    },
    {
      "id": "929942",
      "postDate": "07/15/2020 05:13:43",
      "content": "<p>Experimenting means less waiting.\nSo , i would suggest 5 gpu using one experiment. With this you can also monitor the metrics being calculated and how they are changing while training progress. Resulting in better Understanding.\nElse , your dataset is small , then you can go for individual experiment on each gpu. with small datasets in this way , you can try various different things or methods .</p>",
      "rawMarkdown": "Experimenting means less waiting.\nSo , i would suggest 5 gpu using one experiment. With this you can also monitor the metrics being calculated and how they are changing while training progress. Resulting in better Understanding.\nElse , your dataset is small , then you can go for individual experiment on each gpu. with small datasets in this way , you can try various different things or methods .",
      "votes": null
    },
    {
      "id": "929944",
      "postDate": "07/15/2020 05:16:03",
      "content": "<p>Thanks!</p>",
      "rawMarkdown": "Thanks!",
      "votes": null
    },
    {
      "id": "934428",
      "postDate": "07/18/2020 12:32:09",
      "content": "<p>Can use Pytorch Lightning! Help to deploy multi GPU</p>",
      "rawMarkdown": "Can use Pytorch Lightning! Help to deploy multi GPU",
      "votes": null
    },
    {
      "id": "936309",
      "postDate": "07/20/2020 06:25:28",
      "content": "<p>Thanks your comment!</p>",
      "rawMarkdown": "Thanks your comment!",
      "votes": null
    },
    {
      "id": "936381",
      "postDate": "07/20/2020 07:16:42",
      "content": "<p>I have 4 machines, each with dual GPU so I suppose I could run 8 experiments at one time.\nNot tried 8 but have tried using each machine for a different path.  I only use tensor board to collect data and will occassionaly generate a spread sheet to track things - but my mind not capable of successful multi task when completion of the model takes 8 hours.  It's too easy to look track of the small ideas that can be helpful.</p>\n\n<p>Part is old age (73) and part is my lack of data collection - so I now run 4 versions of an idea on the machines rather than 4 different ideas.  </p>\n\n<p>If you have good data collection - document your path, etc than going in 5 directions at one time will work.  It worked for me from 55 to 65 :)</p>\n\n<p>So I would use all my horses to pull the same wagon to get speed into the system.  Just reading up on potential to cluster my 4 machines using tensor flow so I am hoping to turn the wagon into a jet:)</p>",
      "rawMarkdown": "I have 4 machines, each with dual GPU so I suppose I could run 8 experiments at one time.\nNot tried 8 but have tried using each machine for a different path.  I only use tensor board to collect data and will occassionaly generate a spread sheet to track things - but my mind not capable of successful multi task when completion of the model takes 8 hours.  It's too easy to look track of the small ideas that can be helpful.\n\nPart is old age (73) and part is my lack of data collection - so I now run 4 versions of an idea on the machines rather than 4 different ideas.  \n\nIf you have good data collection - document your path, etc than going in 5 directions at one time will work.  It worked for me from 55 to 65 :)\n\nSo I would use all my horses to pull the same wagon to get speed into the system.  Just reading up on potential to cluster my 4 machines using tensor flow so I am hoping to turn the wagon into a jet:)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 929942,
      "author_name": "prashantarorat",
      "author_url": "",
      "post_date": "07/15/2020 05:13:43",
      "content": "<p>Experimenting means less waiting.\nSo , i would suggest 5 gpu using one experiment. With this you can also monitor the metrics being calculated and how they are changing while training progress. Resulting in better Understanding.\nElse , your dataset is small , then you can go for individual experiment on each gpu. with small datasets in this way , you can try various different things or methods .</p>",
      "votes": null,
      "replies": [
        {
          "id": 929944,
          "author_name": "zxzxs9182",
          "author_url": "",
          "post_date": "07/15/2020 05:16:03",
          "content": "<p>Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 934428,
      "author_name": "doanquanvietnamca",
      "author_url": "",
      "post_date": "07/18/2020 12:32:09",
      "content": "<p>Can use Pytorch Lightning! Help to deploy multi GPU</p>",
      "votes": null,
      "replies": [
        {
          "id": 936309,
          "author_name": "zxzxs9182",
          "author_url": "",
          "post_date": "07/20/2020 06:25:28",
          "content": "<p>Thanks your comment!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 936381,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "07/20/2020 07:16:42",
      "content": "<p>I have 4 machines, each with dual GPU so I suppose I could run 8 experiments at one time.\nNot tried 8 but have tried using each machine for a different path.  I only use tensor board to collect data and will occassionaly generate a spread sheet to track things - but my mind not capable of successful multi task when completion of the model takes 8 hours.  It's too easy to look track of the small ideas that can be helpful.</p>\n\n<p>Part is old age (73) and part is my lack of data collection - so I now run 4 versions of an idea on the machines rather than 4 different ideas.  </p>\n\n<p>If you have good data collection - document your path, etc than going in 5 directions at one time will work.  It worked for me from 55 to 65 :)</p>\n\n<p>So I would use all my horses to pull the same wagon to get speed into the system.  Just reading up on potential to cluster my 4 machines using tensor flow so I am hoping to turn the wagon into a jet:)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "929928": "What is best method to experiment?\n( If I have 5 gpu, \n5 gpu using one experiment\nvs\neach gpu(1) using each experiment)",
    "929942": "Experimenting means less waiting.\nSo , i would suggest 5 gpu using one experiment. With this you can also monitor the metrics being calculated and how they are changing while training progress. Resulting in better Understanding.\nElse , your dataset is small , then you can go for individual experiment on each gpu. with small datasets in this way , you can try various different things or methods .",
    "929944": "Thanks!",
    "934428": "Can use Pytorch Lightning! Help to deploy multi GPU",
    "936309": "Thanks your comment!",
    "936381": "I have 4 machines, each with dual GPU so I suppose I could run 8 experiments at one time.\nNot tried 8 but have tried using each machine for a different path.  I only use tensor board to collect data and will occassionaly generate a spread sheet to track things - but my mind not capable of successful multi task when completion of the model takes 8 hours.  It's too easy to look track of the small ideas that can be helpful.\n\nPart is old age (73) and part is my lack of data collection - so I now run 4 versions of an idea on the machines rather than 4 different ideas.  \n\nIf you have good data collection - document your path, etc than going in 5 directions at one time will work.  It worked for me from 55 to 65 :)\n\nSo I would use all my horses to pull the same wagon to get speed into the system.  Just reading up on potential to cluster my 4 machines using tensor flow so I am hoping to turn the wagon into a jet:)"
  },
  "source": "meta"
}