{
  "id": 105491,
  "title": "EfficientNet not implementing",
  "url": "/competitions/aptos2019-blindness-detection/discussion/105491",
  "author_name": "",
  "post_date": "2019-08-23T13:38:25.644650500Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I m trying to use EfficientNet in my model. I am doing my image preprocessing in different kernel and loading images as numpy array in another kernel, but every time my kernel is going to died just before fit_generator  method of keras ( training code ).  I don't understand please anybody can help...</p>",
  "messages": [
    {
      "id": "606363",
      "postDate": "08/23/2019 13:38:25",
      "content": "<p>I m trying to use EfficientNet in my model. I am doing my image preprocessing in different kernel and loading images as numpy array in another kernel, but every time my kernel is going to died just before fit_generator  method of keras ( training code ).  I don't understand please anybody can help...</p>",
      "rawMarkdown": "I m trying to use EfficientNet in my model. I am doing my image preprocessing in different kernel and loading images as numpy array in another kernel, but every time my kernel is going to died just before fit_generator  method of keras ( training code ).  I don't understand please anybody can help...",
      "votes": null
    },
    {
      "id": "606482",
      "postDate": "08/23/2019 16:35:04",
      "content": "<p>Which EfficientNet are you using ? The processing power increases with the number after the b. B7 from experience requires more processing power than b5 or 6 for example. There are two strategies you can use overcome those limitation during training. First,  use smaller images.  Second you can try smaller batches. </p>\n\n<p>I don't have all the details for what you are trying to do. But, I hope this helps.</p>\n\n<p>Is the kernel dying with out of memory message? </p>",
      "rawMarkdown": "Which EfficientNet are you using ? The processing power increases with the number after the b. B7 from experience requires more processing power than b5 or 6 for example. There are two strategies you can use overcome those limitation during training. First,  use smaller images.  Second you can try smaller batches. \n\nI don't have all the details for what you are trying to do. But, I hope this helps.\n\nIs the kernel dying with out of memory message?",
      "votes": null
    },
    {
      "id": "606513",
      "postDate": "08/23/2019 17:35:08",
      "content": "<p>I am using B5</p>",
      "rawMarkdown": "I am using B5",
      "votes": null
    },
    {
      "id": "606566",
      "postDate": "08/23/2019 18:36:40",
      "content": "<p>Few things I may contribute:\n1. Check if the kernel is getting OOM. My batchsize  had to be relatively low unless ran in fp16. (You could start at 2 for instance and continue increasing it until there seems to be a limit)\n2. Run it local (fine if it's CPU) just to confirm it's not your code. You don't have to actuall train it, but just see that the fit_generator method works</p>\n\n<p>Otherwise see if you can know more from the error messages, can't tell what's the issue without must details :(</p>",
      "rawMarkdown": "Few things I may contribute:\n1. Check if the kernel is getting OOM. My batchsize  had to be relatively low unless ran in fp16. (You could start at 2 for instance and continue increasing it until there seems to be a limit)\n2. Run it local (fine if it's CPU) just to confirm it's not your code. You don't have to actuall train it, but just see that the fit_generator method works\n\nOtherwise see if you can know more from the error messages, can't tell what's the issue without must details :(",
      "votes": null
    },
    {
      "id": "606956",
      "postDate": "08/24/2019 11:19:33",
      "content": "<p>thanks but It won't work...\nplease let me know what type of information you need </p>",
      "rawMarkdown": "thanks but It won't work...\nplease let me know what type of information you need",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 606482,
      "author_name": "abualabed",
      "author_url": "",
      "post_date": "08/23/2019 16:35:04",
      "content": "<p>Which EfficientNet are you using ? The processing power increases with the number after the b. B7 from experience requires more processing power than b5 or 6 for example. There are two strategies you can use overcome those limitation during training. First,  use smaller images.  Second you can try smaller batches. </p>\n\n<p>I don't have all the details for what you are trying to do. But, I hope this helps.</p>\n\n<p>Is the kernel dying with out of memory message? </p>",
      "votes": null,
      "replies": [
        {
          "id": 606513,
          "author_name": "niteshfre",
          "author_url": "",
          "post_date": "08/23/2019 17:35:08",
          "content": "<p>I am using B5</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 606566,
      "author_name": "joonl04",
      "author_url": "",
      "post_date": "08/23/2019 18:36:40",
      "content": "<p>Few things I may contribute:\n1. Check if the kernel is getting OOM. My batchsize  had to be relatively low unless ran in fp16. (You could start at 2 for instance and continue increasing it until there seems to be a limit)\n2. Run it local (fine if it's CPU) just to confirm it's not your code. You don't have to actuall train it, but just see that the fit_generator method works</p>\n\n<p>Otherwise see if you can know more from the error messages, can't tell what's the issue without must details :(</p>",
      "votes": null,
      "replies": [
        {
          "id": 606956,
          "author_name": "niteshfre",
          "author_url": "",
          "post_date": "08/24/2019 11:19:33",
          "content": "<p>thanks but It won't work...\nplease let me know what type of information you need </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "606363": "I m trying to use EfficientNet in my model. I am doing my image preprocessing in different kernel and loading images as numpy array in another kernel, but every time my kernel is going to died just before fit_generator  method of keras ( training code ).  I don't understand please anybody can help...",
    "606482": "Which EfficientNet are you using ? The processing power increases with the number after the b. B7 from experience requires more processing power than b5 or 6 for example. There are two strategies you can use overcome those limitation during training. First,  use smaller images.  Second you can try smaller batches. \n\nI don't have all the details for what you are trying to do. But, I hope this helps.\n\nIs the kernel dying with out of memory message?",
    "606513": "I am using B5",
    "606566": "Few things I may contribute:\n1. Check if the kernel is getting OOM. My batchsize  had to be relatively low unless ran in fp16. (You could start at 2 for instance and continue increasing it until there seems to be a limit)\n2. Run it local (fine if it's CPU) just to confirm it's not your code. You don't have to actuall train it, but just see that the fit_generator method works\n\nOtherwise see if you can know more from the error messages, can't tell what's the issue without must details :(",
    "606956": "thanks but It won't work...\nplease let me know what type of information you need"
  },
  "source": "meta"
}