{
  "id": 77868,
  "title": "How are you doing with martinpiotte 's solution?",
  "url": "/competitions/humpback-whale-identification/discussion/77868",
  "author_name": "",
  "post_date": "2019-01-17T07:33:31.641915300Z",
  "votes": 4,
  "comment_count": 8,
  "views": 0,
  "content": "<p>Hi, how are you guys working with <a href=\"/martinpiotte\">@martinpiotte</a> 's solution? Training from scratch? or load his pretrained model and finetune? Which one gives better performance? \nI am currently training from scratch, but training takes a long time ....</p>",
  "messages": [
    {
      "id": "457292",
      "postDate": "01/17/2019 07:33:31",
      "content": "<p>Hi, how are you guys working with <a href=\"/martinpiotte\">@martinpiotte</a> 's solution? Training from scratch? or load his pretrained model and finetune? Which one gives better performance? \nI am currently training from scratch, but training takes a long time ....</p>",
      "rawMarkdown": "Hi, how are you guys working with @martinpiotte 's solution? Training from scratch? or load his pretrained model and finetune? Which one gives better performance? \nI am currently training from scratch, but training takes a long time ....",
      "votes": null
    },
    {
      "id": "457404",
      "postDate": "01/17/2019 11:01:00",
      "content": "<p>For me its very inestable. Training from scratch LB between 0.812 and 8.50</p>",
      "rawMarkdown": "For me its very inestable. Training from scratch LB between 0.812 and 8.50",
      "votes": null
    },
    {
      "id": "457633",
      "postDate": "01/17/2019 20:38:14",
      "content": "<p>I tried both training from scratch and finetuning. \nUnfortunately, it gets lower score than original pretrained model and I don't know why :(</p>",
      "rawMarkdown": "I tried both training from scratch and finetuning. \nUnfortunately, it gets lower score than original pretrained model and I don't know why :(",
      "votes": null
    },
    {
      "id": "457687",
      "postDate": "01/17/2019 23:58:28",
      "content": "<p>Trained from scratch and did fine tuning with bigger img size</p>",
      "rawMarkdown": "Trained from scratch and did fine tuning with bigger img size",
      "votes": null
    },
    {
      "id": "457720",
      "postDate": "01/18/2019 01:23:37",
      "content": "<p>maybe we need to split validation set, which martinpiotte 's solution does not cover </p>",
      "rawMarkdown": "maybe we need to split validation set, which martinpiotte 's solution does not cover",
      "votes": null
    },
    {
      "id": "457722",
      "postDate": "01/18/2019 01:24:10",
      "content": "<p>good idea, thx</p>",
      "rawMarkdown": "good idea, thx",
      "votes": null
    },
    {
      "id": "457723",
      "postDate": "01/18/2019 01:25:19",
      "content": "<p>have not finish training, yet. But it should not be that case.</p>",
      "rawMarkdown": "have not finish training, yet. But it should not be that case.",
      "votes": null
    },
    {
      "id": "457831",
      "postDate": "01/18/2019 06:53:13",
      "content": "<p>Do you create a validation set? When I try martinpiotte 's solution and triplet loss, I find it's hard to create a validation set.</p>",
      "rawMarkdown": "Do you create a validation set? When I try martinpiotte 's solution and triplet loss, I find it's hard to create a validation set.",
      "votes": null
    },
    {
      "id": "457857",
      "postDate": "01/18/2019 08:06:16",
      "content": "<p>I didn't create the validation, train and use the LB score to check and compare the result. </p>",
      "rawMarkdown": "I didn't create the validation, train and use the LB score to check and compare the result.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 457404,
      "author_name": "maparla",
      "author_url": "",
      "post_date": "01/17/2019 11:01:00",
      "content": "<p>For me its very inestable. Training from scratch LB between 0.812 and 8.50</p>",
      "votes": null,
      "replies": [
        {
          "id": 457720,
          "author_name": "zjucor",
          "author_url": "",
          "post_date": "01/18/2019 01:23:37",
          "content": "<p>maybe we need to split validation set, which martinpiotte 's solution does not cover </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 457633,
      "author_name": "spsancti",
      "author_url": "",
      "post_date": "01/17/2019 20:38:14",
      "content": "<p>I tried both training from scratch and finetuning. \nUnfortunately, it gets lower score than original pretrained model and I don't know why :(</p>",
      "votes": null,
      "replies": [
        {
          "id": 457723,
          "author_name": "zjucor",
          "author_url": "",
          "post_date": "01/18/2019 01:25:19",
          "content": "<p>have not finish training, yet. But it should not be that case.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 457687,
      "author_name": "thuongdinh",
      "author_url": "",
      "post_date": "01/17/2019 23:58:28",
      "content": "<p>Trained from scratch and did fine tuning with bigger img size</p>",
      "votes": null,
      "replies": [
        {
          "id": 457722,
          "author_name": "zjucor",
          "author_url": "",
          "post_date": "01/18/2019 01:24:10",
          "content": "<p>good idea, thx</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 457831,
          "author_name": "smiletat",
          "author_url": "",
          "post_date": "01/18/2019 06:53:13",
          "content": "<p>Do you create a validation set? When I try martinpiotte 's solution and triplet loss, I find it's hard to create a validation set.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 457857,
          "author_name": "thuongdinh",
          "author_url": "",
          "post_date": "01/18/2019 08:06:16",
          "content": "<p>I didn't create the validation, train and use the LB score to check and compare the result. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "457292": "Hi, how are you guys working with @martinpiotte 's solution? Training from scratch? or load his pretrained model and finetune? Which one gives better performance? \nI am currently training from scratch, but training takes a long time ....",
    "457404": "For me its very inestable. Training from scratch LB between 0.812 and 8.50",
    "457633": "I tried both training from scratch and finetuning. \nUnfortunately, it gets lower score than original pretrained model and I don't know why :(",
    "457687": "Trained from scratch and did fine tuning with bigger img size",
    "457720": "maybe we need to split validation set, which martinpiotte 's solution does not cover",
    "457722": "good idea, thx",
    "457723": "have not finish training, yet. But it should not be that case.",
    "457831": "Do you create a validation set? When I try martinpiotte 's solution and triplet loss, I find it's hard to create a validation set.",
    "457857": "I didn't create the validation, train and use the LB score to check and compare the result."
  },
  "source": "meta"
}