{
  "id": 77083,
  "title": "\"Pure\" CNN does not work on whales: looking for suggestions",
  "url": "/competitions/humpback-whale-identification/discussion/77083",
  "author_name": "",
  "post_date": "2019-01-09T10:43:33.095746400Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/fizpok/humpback-whale-2\">This is the link to a kernel</a>\nHi,\nI have created a sample kernel. I can not see errors, but it seems to be overfitting, no mater what I do.\nLooking for suggestions.\nNote: I know I can use pre-trained networks. But I want to understand, why a \"pure\" CNN does not work.</p>\n\n<p>Thanks.</p>",
  "messages": [
    {
      "id": "452907",
      "postDate": "01/09/2019 10:43:33",
      "content": "<p><a href=\"https://www.kaggle.com/fizpok/humpback-whale-2\">This is the link to a kernel</a>\nHi,\nI have created a sample kernel. I can not see errors, but it seems to be overfitting, no mater what I do.\nLooking for suggestions.\nNote: I know I can use pre-trained networks. But I want to understand, why a \"pure\" CNN does not work.</p>\n\n<p>Thanks.</p>",
      "rawMarkdown": "[This is the link to a kernel][1]\n  [1]: https://www.kaggle.com/fizpok/humpback-whale-2\n\nHi,\nI have created a sample kernel. I can not see errors, but it seems to be overfitting, no mater what I do.\nLooking for suggestions.\nNote: I know I can use pre-trained networks. But I want to understand, why a \"pure\" CNN does not work.\n\nThanks.",
      "votes": null
    },
    {
      "id": "452941",
      "postDate": "01/09/2019 11:37:06",
      "content": "<p>I didn't look at the code but you could try increasing regularization. You can let it underfit and then remove regularizations progressively. Also use more data augmentation. Normally it's hard to overfit with modern commonly used techniques</p>",
      "rawMarkdown": "I didn't look at the code but you could try increasing regularization. You can let it underfit and then remove regularizations progressively. Also use more data augmentation. Normally it's hard to overfit with modern commonly used techniques",
      "votes": null
    },
    {
      "id": "453127",
      "postDate": "01/09/2019 18:00:11",
      "content": "<p>If you search for \"one shot learning\" you'll find some resources on why this challenge is hard for normal classification techniques and some potential solutions/workarounds to this problem</p>",
      "rawMarkdown": "If you search for \"one shot learning\" you'll find some resources on why this challenge is hard for normal classification techniques and some potential solutions/workarounds to this problem",
      "votes": null
    },
    {
      "id": "453702",
      "postDate": "01/10/2019 16:27:44",
      "content": "<p>Thread with some good thoughts on the issues:\n<a href=\"https://www.kaggle.com/c/humpback-whale-identification/discussion/74402\">https://www.kaggle.com/c/humpback-whale-identification/discussion/74402</a></p>",
      "rawMarkdown": "Thread with some good thoughts on the issues:\nhttps://www.kaggle.com/c/humpback-whale-identification/discussion/74402",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 452941,
      "author_name": "prajjwal",
      "author_url": "",
      "post_date": "01/09/2019 11:37:06",
      "content": "<p>I didn't look at the code but you could try increasing regularization. You can let it underfit and then remove regularizations progressively. Also use more data augmentation. Normally it's hard to overfit with modern commonly used techniques</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 453127,
      "author_name": "anjum48",
      "author_url": "",
      "post_date": "01/09/2019 18:00:11",
      "content": "<p>If you search for \"one shot learning\" you'll find some resources on why this challenge is hard for normal classification techniques and some potential solutions/workarounds to this problem</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 453702,
      "author_name": "sariabod",
      "author_url": "",
      "post_date": "01/10/2019 16:27:44",
      "content": "<p>Thread with some good thoughts on the issues:\n<a href=\"https://www.kaggle.com/c/humpback-whale-identification/discussion/74402\">https://www.kaggle.com/c/humpback-whale-identification/discussion/74402</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "452907": "[This is the link to a kernel][1]\n  [1]: https://www.kaggle.com/fizpok/humpback-whale-2\n\nHi,\nI have created a sample kernel. I can not see errors, but it seems to be overfitting, no mater what I do.\nLooking for suggestions.\nNote: I know I can use pre-trained networks. But I want to understand, why a \"pure\" CNN does not work.\n\nThanks.",
    "452941": "I didn't look at the code but you could try increasing regularization. You can let it underfit and then remove regularizations progressively. Also use more data augmentation. Normally it's hard to overfit with modern commonly used techniques",
    "453127": "If you search for \"one shot learning\" you'll find some resources on why this challenge is hard for normal classification techniques and some potential solutions/workarounds to this problem",
    "453702": "Thread with some good thoughts on the issues:\nhttps://www.kaggle.com/c/humpback-whale-identification/discussion/74402"
  },
  "source": "meta"
}