{
  "id": 57893,
  "title": "Four images in train data has zero size.",
  "url": "/competitions/avito-demand-prediction/discussion/57893",
  "author_name": "",
  "post_date": "2018-05-30T15:51:45.726676900Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Just notify this.\n<a href=\"https://www.kaggle.com/liujilong/four-image-has-zero-size\">https://www.kaggle.com/liujilong/four-image-has-zero-size</a></p>",
  "messages": [
    {
      "id": "335894",
      "postDate": "05/30/2018 15:51:45",
      "content": "<p>Just notify this.\n<a href=\"https://www.kaggle.com/liujilong/four-image-has-zero-size\">https://www.kaggle.com/liujilong/four-image-has-zero-size</a></p>",
      "rawMarkdown": "Just notify this.\nhttps://www.kaggle.com/liujilong/four-image-has-zero-size",
      "votes": null
    },
    {
      "id": "338546",
      "postDate": "06/05/2018 09:32:19",
      "content": "<p>thanks for sharing ,i am going to train the image data ,how much does the images helps in score may i ask</p>",
      "rawMarkdown": "thanks for sharing ,i am going to train the image data ,how much does the images helps in score may i ask",
      "votes": null
    },
    {
      "id": "338591",
      "postDate": "06/05/2018 11:43:36",
      "content": "<p>Well, It's hard to tell how much image features help. But you can try different approaches by yourself and see whether or not.</p>\n\n<p>Three approaches for example:</p>\n\n<ul>\n<li>Use image meta features (width, height, brightness, saturation, colorful....). There are some great kernels you can refer to.</li>\n<li>Use image scores. Max probability of label predicted by Image Net, or Google's NIMA score.</li>\n<li>Feed raw image pixel in CNN models.</li>\n</ul>\n\n<p>For me, All three approaches helped.</p>",
      "rawMarkdown": "Well, It's hard to tell how much image features help. But you can try different approaches by yourself and see whether or not.\n\nThree approaches for example:\n\n - Use image meta features (width, height, brightness, saturation, colorful....). There are some great kernels you can refer to.\n - Use image scores. Max probability of label predicted by Image Net, or Google's NIMA score.\n - Feed raw image pixel in CNN models.\n\n\n\nFor me, All three approaches helped.",
      "votes": null
    },
    {
      "id": "338609",
      "postDate": "06/05/2018 12:34:04",
      "content": "<p>Thanks for that,I will try,image_top_1 is an important feature,I think it will help alot .I get only 0.2216 using just the csv data,what about you?</p>",
      "rawMarkdown": "Thanks for that,I will try,image_top_1 is an important feature,I think it will help alot .I get only 0.2216 using just the csv data,what about you?",
      "votes": null
    },
    {
      "id": "338628",
      "postDate": "06/05/2018 13:18:42",
      "content": "<p>Can you kindly explain how image_top_1 is important? I can't seem to.</p>",
      "rawMarkdown": "Can you kindly explain how image_top_1 is important? I can't seem to.",
      "votes": null
    },
    {
      "id": "338637",
      "postDate": "06/05/2018 13:30:24",
      "content": "<p>just theLGB result 。list the importance for each feature,imag_top_1 has high score</p>",
      "rawMarkdown": "just theLGB result 。list the importance for each feature,imag_top_1 has high score",
      "votes": null
    },
    {
      "id": "338651",
      "postDate": "06/05/2018 14:04:55",
      "content": "<p>LGB:  0.2205  </p>\n\n<p>NN: 0.2194  </p>\n\n<p>all 5 fold, use all features.</p>",
      "rawMarkdown": "LGB:  0.2205  \n\nNN: 0.2194  \n\nall 5 fold, use all features.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 338546,
      "author_name": "eangle12138",
      "author_url": "",
      "post_date": "06/05/2018 09:32:19",
      "content": "<p>thanks for sharing ,i am going to train the image data ,how much does the images helps in score may i ask</p>",
      "votes": null,
      "replies": [
        {
          "id": 338591,
          "author_name": "liujilong",
          "author_url": "",
          "post_date": "06/05/2018 11:43:36",
          "content": "<p>Well, It's hard to tell how much image features help. But you can try different approaches by yourself and see whether or not.</p>\n\n<p>Three approaches for example:</p>\n\n<ul>\n<li>Use image meta features (width, height, brightness, saturation, colorful....). There are some great kernels you can refer to.</li>\n<li>Use image scores. Max probability of label predicted by Image Net, or Google's NIMA score.</li>\n<li>Feed raw image pixel in CNN models.</li>\n</ul>\n\n<p>For me, All three approaches helped.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 338609,
          "author_name": "eangle12138",
          "author_url": "",
          "post_date": "06/05/2018 12:34:04",
          "content": "<p>Thanks for that,I will try,image_top_1 is an important feature,I think it will help alot .I get only 0.2216 using just the csv data,what about you?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 338628,
          "author_name": "temibabs",
          "author_url": "",
          "post_date": "06/05/2018 13:18:42",
          "content": "<p>Can you kindly explain how image_top_1 is important? I can't seem to.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 338637,
          "author_name": "eangle12138",
          "author_url": "",
          "post_date": "06/05/2018 13:30:24",
          "content": "<p>just theLGB result 。list the importance for each feature,imag_top_1 has high score</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 338651,
          "author_name": "liujilong",
          "author_url": "",
          "post_date": "06/05/2018 14:04:55",
          "content": "<p>LGB:  0.2205  </p>\n\n<p>NN: 0.2194  </p>\n\n<p>all 5 fold, use all features.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "335894": "Just notify this.\nhttps://www.kaggle.com/liujilong/four-image-has-zero-size",
    "338546": "thanks for sharing ,i am going to train the image data ,how much does the images helps in score may i ask",
    "338591": "Well, It's hard to tell how much image features help. But you can try different approaches by yourself and see whether or not.\n\nThree approaches for example:\n\n - Use image meta features (width, height, brightness, saturation, colorful....). There are some great kernels you can refer to.\n - Use image scores. Max probability of label predicted by Image Net, or Google's NIMA score.\n - Feed raw image pixel in CNN models.\n\n\n\nFor me, All three approaches helped.",
    "338609": "Thanks for that,I will try,image_top_1 is an important feature,I think it will help alot .I get only 0.2216 using just the csv data,what about you?",
    "338628": "Can you kindly explain how image_top_1 is important? I can't seem to.",
    "338637": "just theLGB result 。list the importance for each feature,imag_top_1 has high score",
    "338651": "LGB:  0.2205  \n\nNN: 0.2194  \n\nall 5 fold, use all features."
  },
  "source": "meta"
}