{
  "id": 79582,
  "title": "Anyone worried their kernel might fail/error?",
  "url": "/competitions/quora-insincere-questions-classification/discussion/79582",
  "author_name": "",
  "post_date": "2019-02-05T18:07:06.825046700Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>It is so hard to write error-free code. Don't know if the test data might break the kernel. Keeping fingers crossed</p>",
  "messages": [
    {
      "id": "466631",
      "postDate": "02/05/2019 18:07:06",
      "content": "<p>It is so hard to write error-free code. Don't know if the test data might break the kernel. Keeping fingers crossed</p>",
      "rawMarkdown": "It is so hard to write error-free code. Don't know if the test data might break the kernel. Keeping fingers crossed",
      "votes": null
    },
    {
      "id": "466656",
      "postDate": "02/05/2019 19:17:06",
      "content": "<p>There are way too many things I'm worried about right now (including this) ! It has been 3 months of hard work. In India, it's close to 1 a.m. at the time of typing. I'm so anxious with just 5 hours to go.</p>",
      "rawMarkdown": "There are way too many things I'm worried about right now (including this) ! It has been 3 months of hard work. In India, it's close to 1 a.m. at the time of typing. I'm so anxious with just 5 hours to go.",
      "votes": null
    },
    {
      "id": "466664",
      "postDate": "02/05/2019 19:29:33",
      "content": "<p>I just changed the batch size in my submission kernel and it broke. So many hardcoded things in models and kernels. </p>",
      "rawMarkdown": "I just changed the batch size in my submission kernel and it broke. So many hardcoded things in models and kernels.",
      "votes": null
    },
    {
      "id": "466713",
      "postDate": "02/05/2019 21:02:37",
      "content": "<h3>Here's some tricks I learned in Mercari Price Suggestion Challenge:</h3>\n\n<ol>\n<li>Add try-except in your preprocessing functions.</li>\n<li>Do <code>pandas.DataFrame.fillna()</code> since we don't know if there are null data in stage 2. </li>\n<li>(Optional) Once an object is no longer used, delete it and do <code>gc.collect()</code> to prevent OOM.</li>\n<li>(Optional) Delete the model and refresh CUDA memory at the end of every fold for extra stability.</li>\n</ol>\n\n<h3>Make sure your kernel can survive these:</h3>\n\n<ol>\n<li>Repeat your test data to 8x size. (Average word count could be larger in stage 2, so use 8x instead of 7x)</li>\n<li>Manually add some null data in both your train data and test data.</li>\n<li>Your kernel should complete in 6840 seconds (95%) in case you get a slow kernel in stage 2.</li>\n</ol>",
      "rawMarkdown": "### Here's some tricks I learned in Mercari Price Suggestion Challenge:    \n1. Add try-except in your preprocessing functions.\n2. Do `pandas.DataFrame.fillna()` since we don't know if there are null data in stage 2. \n3. (Optional) Once an object is no longer used, delete it and do `gc.collect()` to prevent OOM.\n4. (Optional) Delete the model and refresh CUDA memory at the end of every fold for extra stability.\n\n### Make sure your kernel can survive these: \n1. Repeat your test data to 8x size. (Average word count could be larger in stage 2, so use 8x instead of 7x)\n2. Manually add some null data in both your train data and test data.\n3. Your kernel should complete in 6840 seconds (95%) in case you get a slow kernel in stage 2.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 466656,
      "author_name": "tarunpaparaju",
      "author_url": "",
      "post_date": "02/05/2019 19:17:06",
      "content": "<p>There are way too many things I'm worried about right now (including this) ! It has been 3 months of hard work. In India, it's close to 1 a.m. at the time of typing. I'm so anxious with just 5 hours to go.</p>",
      "votes": null,
      "replies": [
        {
          "id": 466664,
          "author_name": "mlwhiz",
          "author_url": "",
          "post_date": "02/05/2019 19:29:33",
          "content": "<p>I just changed the batch size in my submission kernel and it broke. So many hardcoded things in models and kernels. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 466713,
      "author_name": "jerrykuo7727",
      "author_url": "",
      "post_date": "02/05/2019 21:02:37",
      "content": "<h3>Here's some tricks I learned in Mercari Price Suggestion Challenge:</h3>\n\n<ol>\n<li>Add try-except in your preprocessing functions.</li>\n<li>Do <code>pandas.DataFrame.fillna()</code> since we don't know if there are null data in stage 2. </li>\n<li>(Optional) Once an object is no longer used, delete it and do <code>gc.collect()</code> to prevent OOM.</li>\n<li>(Optional) Delete the model and refresh CUDA memory at the end of every fold for extra stability.</li>\n</ol>\n\n<h3>Make sure your kernel can survive these:</h3>\n\n<ol>\n<li>Repeat your test data to 8x size. (Average word count could be larger in stage 2, so use 8x instead of 7x)</li>\n<li>Manually add some null data in both your train data and test data.</li>\n<li>Your kernel should complete in 6840 seconds (95%) in case you get a slow kernel in stage 2.</li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "466631": "It is so hard to write error-free code. Don't know if the test data might break the kernel. Keeping fingers crossed",
    "466656": "There are way too many things I'm worried about right now (including this) ! It has been 3 months of hard work. In India, it's close to 1 a.m. at the time of typing. I'm so anxious with just 5 hours to go.",
    "466664": "I just changed the batch size in my submission kernel and it broke. So many hardcoded things in models and kernels.",
    "466713": "### Here's some tricks I learned in Mercari Price Suggestion Challenge:    \n1. Add try-except in your preprocessing functions.\n2. Do `pandas.DataFrame.fillna()` since we don't know if there are null data in stage 2. \n3. (Optional) Once an object is no longer used, delete it and do `gc.collect()` to prevent OOM.\n4. (Optional) Delete the model and refresh CUDA memory at the end of every fold for extra stability.\n\n### Make sure your kernel can survive these: \n1. Repeat your test data to 8x size. (Average word count could be larger in stage 2, so use 8x instead of 7x)\n2. Manually add some null data in both your train data and test data.\n3. Your kernel should complete in 6840 seconds (95%) in case you get a slow kernel in stage 2."
  },
  "source": "meta"
}