{
  "id": 79787,
  "title": "Any tips for starting playing a CV competition?",
  "url": "/competitions/humpback-whale-identification/discussion/79787",
  "author_name": "",
  "post_date": "2019-02-07T13:47:11.935943100Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I've been staying away from CV competitions mainly because of the following reason : \n1. High hardware requirement :  My graphics card (GTX1060)  just can't load big models. Thus I have to use kernels. Which leads me to the second problem.\n2. Environment setting is hard :  I have problems replicating top results. They have the pretrained weights in the local machine where I don't have them in the kernels. This one can be evaded since I can upload them to datasets then load them back in. But it's still a bit painful.</p>\n\n<p>And also would you guys share your hardware specs? Are there any additional tips to play CV competitions? \nThanks in advance :)</p>",
  "messages": [
    {
      "id": "467652",
      "postDate": "02/07/2019 13:47:11",
      "content": "<p>I've been staying away from CV competitions mainly because of the following reason : \n1. High hardware requirement :  My graphics card (GTX1060)  just can't load big models. Thus I have to use kernels. Which leads me to the second problem.\n2. Environment setting is hard :  I have problems replicating top results. They have the pretrained weights in the local machine where I don't have them in the kernels. This one can be evaded since I can upload them to datasets then load them back in. But it's still a bit painful.</p>\n\n<p>And also would you guys share your hardware specs? Are there any additional tips to play CV competitions? \nThanks in advance :)</p>",
      "rawMarkdown": "I've been staying away from CV competitions mainly because of the following reason : \n1. High hardware requirement :  My graphics card (GTX1060)  just can't load big models. Thus I have to use kernels. Which leads me to the second problem.\n2. Environment setting is hard :  I have problems replicating top results. They have the pretrained weights in the local machine where I don't have them in the kernels. This one can be evaded since I can upload them to datasets then load them back in. But it's still a bit painful.\n\nAnd also would you guys share your hardware specs? Are there any additional tips to play CV competitions? \nThanks in advance :)",
      "votes": null
    },
    {
      "id": "467662",
      "postDate": "02/07/2019 14:16:25",
      "content": "<p>My 2 cents:\n1. enter your first competition as soon as possible\n2. do not care about the LB but about what you can learn from it. \n3. read discussions and kernels\n4. check solutions from relevant competitions\n5. play around\n6. have fun</p>\n\n<p>BTW: kernels only will not give you a medal here but you can learn a bunch. And pretrained weights are available.</p>",
      "rawMarkdown": "My 2 cents:\n1. enter your first competition as soon as possible\n2. do not care about the LB but about what you can learn from it. \n3. read discussions and kernels\n4. check solutions from relevant competitions\n5. play around\n6. have fun\n\nBTW: kernels only will not give you a medal here but you can learn a bunch. And pretrained weights are available.",
      "votes": null
    },
    {
      "id": "467668",
      "postDate": "02/07/2019 14:31:14",
      "content": "<p>Thanks for the advice! Looking at the kernel the specs aren't very bad. It's just inconvenient to use. May I ask what specs do you have?</p>",
      "rawMarkdown": "Thanks for the advice! Looking at the kernel the specs aren't very bad. It's just inconvenient to use. May I ask what specs do you have?",
      "votes": null
    },
    {
      "id": "467674",
      "postDate": "02/07/2019 14:47:59",
      "content": "<p>i always use kernels. In this comp I forked <a href=\"https://www.kaggle.com/iafoss/similarity-resnext50-0-740-lb-kernel-time-limit\">https://www.kaggle.com/iafoss/similarity-resnext50-0-740-lb-kernel-time-limit</a> . I was lucky with TGS to win a medal (tiny dataset) but who cares, I am learning in this one too :)</p>",
      "rawMarkdown": "i always use kernels. In this comp I forked https://www.kaggle.com/iafoss/similarity-resnext50-0-740-lb-kernel-time-limit . I was lucky with TGS to win a medal (tiny dataset) but who cares, I am learning in this one too :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 467662,
      "author_name": "valanm",
      "author_url": "",
      "post_date": "02/07/2019 14:16:25",
      "content": "<p>My 2 cents:\n1. enter your first competition as soon as possible\n2. do not care about the LB but about what you can learn from it. \n3. read discussions and kernels\n4. check solutions from relevant competitions\n5. play around\n6. have fun</p>\n\n<p>BTW: kernels only will not give you a medal here but you can learn a bunch. And pretrained weights are available.</p>",
      "votes": null,
      "replies": [
        {
          "id": 467668,
          "author_name": "icemekaveli",
          "author_url": "",
          "post_date": "02/07/2019 14:31:14",
          "content": "<p>Thanks for the advice! Looking at the kernel the specs aren't very bad. It's just inconvenient to use. May I ask what specs do you have?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 467674,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "02/07/2019 14:47:59",
          "content": "<p>i always use kernels. In this comp I forked <a href=\"https://www.kaggle.com/iafoss/similarity-resnext50-0-740-lb-kernel-time-limit\">https://www.kaggle.com/iafoss/similarity-resnext50-0-740-lb-kernel-time-limit</a> . I was lucky with TGS to win a medal (tiny dataset) but who cares, I am learning in this one too :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "467652": "I've been staying away from CV competitions mainly because of the following reason : \n1. High hardware requirement :  My graphics card (GTX1060)  just can't load big models. Thus I have to use kernels. Which leads me to the second problem.\n2. Environment setting is hard :  I have problems replicating top results. They have the pretrained weights in the local machine where I don't have them in the kernels. This one can be evaded since I can upload them to datasets then load them back in. But it's still a bit painful.\n\nAnd also would you guys share your hardware specs? Are there any additional tips to play CV competitions? \nThanks in advance :)",
    "467662": "My 2 cents:\n1. enter your first competition as soon as possible\n2. do not care about the LB but about what you can learn from it. \n3. read discussions and kernels\n4. check solutions from relevant competitions\n5. play around\n6. have fun\n\nBTW: kernels only will not give you a medal here but you can learn a bunch. And pretrained weights are available.",
    "467668": "Thanks for the advice! Looking at the kernel the specs aren't very bad. It's just inconvenient to use. May I ask what specs do you have?",
    "467674": "i always use kernels. In this comp I forked https://www.kaggle.com/iafoss/similarity-resnext50-0-740-lb-kernel-time-limit . I was lucky with TGS to win a medal (tiny dataset) but who cares, I am learning in this one too :)"
  },
  "source": "meta"
}