{
  "id": 234434,
  "title": "STEPS_PER_EXECUTIONS",
  "url": "/competitions/herbarium-2021-fgvc8/discussion/234434",
  "author_name": "Luigi Saetta",
  "post_date": "2021-04-24T07:53:04.799000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Sometimes simply copy someone else NB and changing can seem a good way to start. And sharing Dataset is always done to help someone else, and increase the participation to a competition.<br>\nBut we should be careful in what we use&amp;copy. <br>\nThere is a Notebook that has been published using the herb2021-256 dataset that contains many things that made me think (and it is always good to think). </p>\n<p>One was the STEPS_PER_EXECUTIONS parameters</p>\n<p>well, my suggestion is NOT to use unless you don't do a careful examination of what actually put there. <br>\nThe value showed in the Notebook slowed my NB by 3x. Simply the choice: just don't.<br>\nIf you have no idea, the default is the best idea.</p>\n<p>Another good practice would be to vote and cite if you're using someone else code&amp;thing… but this is another story.</p>\n<p>At the end, the goal should be learn and have fun…. </p>",
  "messages": [
    {
      "id": 1284330,
      "postDate": "2021-04-25T19:42:48.840Z",
      "content": "<p>Agreed. I forgot to include reference to Martin Goerner's notebook \"Getting started with 100+ flowers on TPU\". I've fixed that, and I also published an <a href=\"https://www.kaggle.com/atamazian/herb-2021-tfrecords-effnet-inference-debug\" target=\"_blank\">inference notebook </a>. Unfortunately, it doesn't work due to memory overflow.</p>",
      "rawMarkdown": "Agreed. I forgot to include reference to Martin Goerner's notebook \"Getting started with 100+ flowers on TPU\". I've fixed that, and I also published an [inference notebook ](https://www.kaggle.com/atamazian/herb-2021-tfrecords-effnet-inference-debug). Unfortunately, it doesn't work due to memory overflow.",
      "votes": 1,
      "replies": [
        {
          "id": 1284831,
          "postDate": "2021-04-26T09:53:13.657Z",
          "content": "<p>Inference is not easy since the test set is also big. The only way I have found it to do the predict on batches. That means that for every tfrecord file I need to iterate over batches… </p>",
          "rawMarkdown": "Inference is not easy since the test set is also big. The only way I have found it to do the predict on batches. That means that for every tfrecord file I need to iterate over batches... \n",
          "votes": 1
        },
        {
          "id": 1284956,
          "postDate": "2021-04-26T12:58:59.547Z",
          "content": "<p>I followed you advice, and it's working! It takes ~12 min to complete inference on GPU.</p>",
          "rawMarkdown": "I followed you advice, and it's working! It takes ~12 min to complete inference on GPU.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1282729,
      "postDate": "2021-04-24T07:53:04.800Z",
      "content": "<p>Sometimes simply copy someone else NB and changing can seem a good way to start. And sharing Dataset is always done to help someone else, and increase the participation to a competition.<br>\nBut we should be careful in what we use&amp;copy. <br>\nThere is a Notebook that has been published using the herb2021-256 dataset that contains many things that made me think (and it is always good to think). </p>\n<p>One was the STEPS_PER_EXECUTIONS parameters</p>\n<p>well, my suggestion is NOT to use unless you don't do a careful examination of what actually put there. <br>\nThe value showed in the Notebook slowed my NB by 3x. Simply the choice: just don't.<br>\nIf you have no idea, the default is the best idea.</p>\n<p>Another good practice would be to vote and cite if you're using someone else code&amp;thing… but this is another story.</p>\n<p>At the end, the goal should be learn and have fun…. </p>",
      "rawMarkdown": "Sometimes simply copy someone else NB and changing can seem a good way to start. And sharing Dataset is always done to help someone else, and increase the participation to a competition.\nBut we should be careful in what we use&copy. \nThere is a Notebook that has been published using the herb2021-256 dataset that contains many things that made me think (and it is always good to think). \n\nOne was the STEPS_PER_EXECUTIONS parameters\n\nwell, my suggestion is NOT to use unless you don't do a careful examination of what actually put there. \nThe value showed in the Notebook slowed my NB by 3x. Simply the choice: just don't.\nIf you have no idea, the default is the best idea.\n\nAnother good practice would be to vote and cite if you're using someone else code&thing... but this is another story.\n\nAt the end, the goal should be learn and have fun.... ",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 1284330,
      "author_name": "Araik Tamazian",
      "author_url": "",
      "post_date": "2021-04-25T19:42:48.840000",
      "content": "<p>Agreed. I forgot to include reference to Martin Goerner's notebook \"Getting started with 100+ flowers on TPU\". I've fixed that, and I also published an <a href=\"https://www.kaggle.com/atamazian/herb-2021-tfrecords-effnet-inference-debug\" target=\"_blank\">inference notebook </a>. Unfortunately, it doesn't work due to memory overflow.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1284831,
          "author_name": "Luigi Saetta",
          "author_url": "",
          "post_date": "2021-04-26T09:53:13.657000",
          "content": "<p>Inference is not easy since the test set is also big. The only way I have found it to do the predict on batches. That means that for every tfrecord file I need to iterate over batches… </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1284956,
          "author_name": "Araik Tamazian",
          "author_url": "",
          "post_date": "2021-04-26T12:58:59.547000",
          "content": "<p>I followed you advice, and it's working! It takes ~12 min to complete inference on GPU.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1284330": "Agreed. I forgot to include reference to Martin Goerner's notebook \"Getting started with 100+ flowers on TPU\". I've fixed that, and I also published an [inference notebook ](https://www.kaggle.com/atamazian/herb-2021-tfrecords-effnet-inference-debug). Unfortunately, it doesn't work due to memory overflow.",
    "1282729": "Sometimes simply copy someone else NB and changing can seem a good way to start. And sharing Dataset is always done to help someone else, and increase the participation to a competition.\nBut we should be careful in what we use&copy. \nThere is a Notebook that has been published using the herb2021-256 dataset that contains many things that made me think (and it is always good to think). \n\nOne was the STEPS_PER_EXECUTIONS parameters\n\nwell, my suggestion is NOT to use unless you don't do a careful examination of what actually put there. \nThe value showed in the Notebook slowed my NB by 3x. Simply the choice: just don't.\nIf you have no idea, the default is the best idea.\n\nAnother good practice would be to vote and cite if you're using someone else code&thing... but this is another story.\n\nAt the end, the goal should be learn and have fun.... "
  }
}