{
  "id": 71724,
  "title": "Time to train a single epoch is more than 24 hr ...",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/71724",
  "author_name": "Peterzhang",
  "post_date": "2018-11-16T01:54:43.667000",
  "votes": 2,
  "comment_count": 11,
  "views": 0,
  "content": "<p>I use resnet50, 224x224x3 and 3x 1080ti, what about your guys? </p>",
  "messages": [
    {
      "id": 422267,
      "postDate": "2018-11-16T01:54:43.667Z",
      "content": "<p>I use resnet50, 224x224x3 and 3x 1080ti, what about your guys? </p>",
      "rawMarkdown": "I use resnet50, 224x224x3 and 3x 1080ti, what about your guys? ",
      "votes": 2
    },
    {
      "id": 422302,
      "postDate": "2018-11-16T03:27:33.300Z",
      "content": "<p>I think the speed is actually normal, as you are using 224 images</p>",
      "rawMarkdown": "I think the speed is actually normal, as you are using 224 images",
      "replies": [
        {
          "id": 422448,
          "postDate": "2018-11-16T08:31:47.777Z",
          "content": "<p>Seems realistic. I trained 128x128 ResNet50 on P100 for a day.</p>",
          "rawMarkdown": "Seems realistic. I trained 128x128 ResNet50 on P100 for a day.",
          "votes": 1
        },
        {
          "id": 422466,
          "postDate": "2018-11-16T08:59:20.817Z",
          "content": "<p><a href=\"/beluga\">@beluga</a> how many epochs did you train that one day model for?</p>",
          "rawMarkdown": "@beluga how many epochs did you train that one day model for?"
        },
        {
          "id": 422515,
          "postDate": "2018-11-16T10:32:10.510Z",
          "content": "<p>This is computable. The data set contains about 50 million pictures. I think even 24hours is not enough to train a complete epoch.</p>",
          "rawMarkdown": "This is computable. The data set contains about 50 million pictures. I think even 24hours is not enough to train a complete epoch.",
          "votes": 1
        },
        {
          "id": 422519,
          "postDate": "2018-11-16T10:39:06.733Z",
          "content": "<p><a href=\"/beluga\">@beluga</a> how many epochs did you train that one day model for?\nI it was actually  0.8 epoch. I use preemptible GPUs they don't live for more than 24 h :(</p>",
          "rawMarkdown": "@beluga how many epochs did you train that one day model for?\nI it was actually  0.8 epoch. I use preemptible GPUs they don't live for more than 24 h :(",
          "votes": 1
        },
        {
          "id": 422523,
          "postDate": "2018-11-16T10:53:07.723Z",
          "content": "<p>Thanks <a href=\"/beluga\">@beluga</a>,  and @Gary-DeepLearning 's point is why I asked. I have not used the whole data yet because of the time limit <a href=\"/beluga\">@beluga</a> mentioned.</p>",
          "rawMarkdown": "Thanks @beluga,  and @Gary-DeepLearning 's point is why I asked. I have not used the whole data yet because of the time limit @beluga mentioned."
        },
        {
          "id": 422612,
          "postDate": "2018-11-16T13:59:46.090Z",
          "content": "<p>Hi, could you tell me how to @ someone ? </p>",
          "rawMarkdown": "Hi, could you tell me how to @ someone ? "
        },
        {
          "id": 422660,
          "postDate": "2018-11-16T15:29:39.913Z",
          "content": "<p>Just type <a href=\"/tangss\">@tangss</a></p>",
          "rawMarkdown": "Just type @tangss"
        },
        {
          "id": 422721,
          "postDate": "2018-11-16T17:06:10.923Z",
          "content": "<p>Resnet34, 256x256x1, batchsize=48, full dataset (99 csv for train 1 for validation (from shuffle csv kernel)); Titan X: 52h per epoch (with evalution after each csv, without 40h per epoch)</p>",
          "rawMarkdown": "Resnet34, 256x256x1, batchsize=48, full dataset (99 csv for train 1 for validation (from shuffle csv kernel)); Titan X: 52h per epoch (with evalution after each csv, without 40h per epoch)",
          "votes": 1
        },
        {
          "id": 423054,
          "postDate": "2018-11-17T11:05:26.933Z",
          "content": "<p>I got it, thanks. let me try <a href=\"/beluga\">@beluga</a></p>",
          "rawMarkdown": "I got it, thanks. let me try @beluga"
        }
      ]
    },
    {
      "id": 422284,
      "postDate": "2018-11-16T02:34:44.553Z",
      "content": "<p>How to generate data on training, I guess the IO maybe so slow in your model.\nOr you can load the pre_trained model trained on ImageNet and fine tune on the Doodle dataset. </p>",
      "rawMarkdown": "How to generate data on training, I guess the IO maybe so slow in your model.\nOr you can load the pre_trained model trained on ImageNet and fine tune on the Doodle dataset. "
    }
  ],
  "comments": [
    {
      "id": 422302,
      "author_name": "Strideradu",
      "author_url": "",
      "post_date": "2018-11-16T03:27:33.300000",
      "content": "<p>I think the speed is actually normal, as you are using 224 images</p>",
      "votes": 0,
      "replies": [
        {
          "id": 422448,
          "author_name": "beluga",
          "author_url": "",
          "post_date": "2018-11-16T08:31:47.777000",
          "content": "<p>Seems realistic. I trained 128x128 ResNet50 on P100 for a day.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 422466,
          "author_name": "YaGana Sheriff-Hussaini",
          "author_url": "",
          "post_date": "2018-11-16T08:59:20.817000",
          "content": "<p><a href=\"/beluga\">@beluga</a> how many epochs did you train that one day model for?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 422515,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2018-11-16T10:32:10.510000",
          "content": "<p>This is computable. The data set contains about 50 million pictures. I think even 24hours is not enough to train a complete epoch.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 422519,
          "author_name": "beluga",
          "author_url": "",
          "post_date": "2018-11-16T10:39:06.733000",
          "content": "<p><a href=\"/beluga\">@beluga</a> how many epochs did you train that one day model for?\nI it was actually  0.8 epoch. I use preemptible GPUs they don't live for more than 24 h :(</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 422523,
          "author_name": "YaGana Sheriff-Hussaini",
          "author_url": "",
          "post_date": "2018-11-16T10:53:07.723000",
          "content": "<p>Thanks <a href=\"/beluga\">@beluga</a>,  and @Gary-DeepLearning 's point is why I asked. I have not used the whole data yet because of the time limit <a href=\"/beluga\">@beluga</a> mentioned.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 422612,
          "author_name": "upup",
          "author_url": "",
          "post_date": "2018-11-16T13:59:46.090000",
          "content": "<p>Hi, could you tell me how to @ someone ? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 422660,
          "author_name": "beluga",
          "author_url": "",
          "post_date": "2018-11-16T15:29:39.913000",
          "content": "<p>Just type <a href=\"/tangss\">@tangss</a></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 422721,
          "author_name": "Valentyn Vaityshyn",
          "author_url": "",
          "post_date": "2018-11-16T17:06:10.923000",
          "content": "<p>Resnet34, 256x256x1, batchsize=48, full dataset (99 csv for train 1 for validation (from shuffle csv kernel)); Titan X: 52h per epoch (with evalution after each csv, without 40h per epoch)</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 423054,
          "author_name": "upup",
          "author_url": "",
          "post_date": "2018-11-17T11:05:26.933000",
          "content": "<p>I got it, thanks. let me try <a href=\"/beluga\">@beluga</a></p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 422284,
      "author_name": "upup",
      "author_url": "",
      "post_date": "2018-11-16T02:34:44.553000",
      "content": "<p>How to generate data on training, I guess the IO maybe so slow in your model.\nOr you can load the pre_trained model trained on ImageNet and fine tune on the Doodle dataset. </p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "422267": "I use resnet50, 224x224x3 and 3x 1080ti, what about your guys? ",
    "422302": "I think the speed is actually normal, as you are using 224 images",
    "422284": "How to generate data on training, I guess the IO maybe so slow in your model.\nOr you can load the pre_trained model trained on ImageNet and fine tune on the Doodle dataset. "
  }
}