{
  "id": 234336,
  "title": "GPU timings",
  "url": "/competitions/plant-pathology-2021-fgvc8/discussion/234336",
  "author_name": "Poojan Patel",
  "post_date": "2021-04-23T18:25:31.556000",
  "votes": 0,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hello,<br>\nI am using keras, pretrained VGG19 and ImageDataGenerator with flow_from_dataframe method. No extra preprocessing. Still it is taking almost 50 min to complete an epoch. Now notebooks which has GPU time &lt;= 2 hrs are accepted, then how can i submit my notebook?</p>",
  "messages": [
    {
      "id": 1282555,
      "postDate": "2021-04-24T04:24:24.543Z",
      "content": "<p>I think GPU is not used optimally, You can adjust batch size and increase max_queue_size to increase the use of GPU. In my case after adjusting these parameters, one epoch is completed in 10 minutes</p>",
      "rawMarkdown": "I think GPU is not used optimally, You can adjust batch size and increase max_queue_size to increase the use of GPU. In my case after adjusting these parameters, one epoch is completed in 10 minutes",
      "votes": 1
    },
    {
      "id": 1282640,
      "postDate": "2021-04-24T06:27:14.290Z",
      "content": "<p>Try this approach for submission.</p>\n<ul>\n<li>Train your model and save it as an h5 file. Make sure to \"Save output\" while committing your training notebook</li>\n<li>Create a new notebook for submission. Import the output of the training notebook as a dataset, and load the model into your editor</li>\n<li>Use the loaded model to make predictions. </li>\n</ul>",
      "rawMarkdown": "Try this approach for submission.\n- Train your model and save it as an h5 file. Make sure to \"Save output\" while committing your training notebook\n- Create a new notebook for submission. Import the output of the training notebook as a dataset, and load the model into your editor\n- Use the loaded model to make predictions. ",
      "votes": 2
    },
    {
      "id": 1284221,
      "postDate": "2021-04-25T17:42:50.780Z",
      "content": "<p>You can speed up computations if you switch to Tensorflow and use TFRecords with augmentations applied beforehand. Check out <a href=\"https://www.kaggle.com/nickuzmenkov/pp2021-tpu-tf-inference\" target=\"_blank\">this notebook</a>.</p>",
      "rawMarkdown": "You can speed up computations if you switch to Tensorflow and use TFRecords with augmentations applied beforehand. Check out [this notebook](https://www.kaggle.com/nickuzmenkov/pp2021-tpu-tf-inference)."
    },
    {
      "id": 1282268,
      "postDate": "2021-04-23T18:25:31.557Z",
      "content": "<p>Hello,<br>\nI am using keras, pretrained VGG19 and ImageDataGenerator with flow_from_dataframe method. No extra preprocessing. Still it is taking almost 50 min to complete an epoch. Now notebooks which has GPU time &lt;= 2 hrs are accepted, then how can i submit my notebook?</p>",
      "rawMarkdown": "Hello,\nI am using keras, pretrained VGG19 and ImageDataGenerator with flow_from_dataframe method. No extra preprocessing. Still it is taking almost 50 min to complete an epoch. Now notebooks which has GPU time <= 2 hrs are accepted, then how can i submit my notebook?"
    },
    {
      "id": 1282590,
      "postDate": "2021-04-24T05:20:05.227Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 1283902,
      "postDate": "2021-04-25T11:43:38.663Z",
      "content": "<p><a href=\"https://www.kaggle.com/mreenav\" target=\"_blank\">@mreenav</a> Thank you for Nice Idea!</p>",
      "rawMarkdown": "@mreenav Thank you for Nice Idea!"
    }
  ],
  "comments": [
    {
      "id": 1282555,
      "author_name": "Rajat Sanjay Patel",
      "author_url": "",
      "post_date": "2021-04-24T04:24:24.543000",
      "content": "<p>I think GPU is not used optimally, You can adjust batch size and increase max_queue_size to increase the use of GPU. In my case after adjusting these parameters, one epoch is completed in 10 minutes</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1282640,
      "author_name": "Mreenav S. Deka",
      "author_url": "",
      "post_date": "2021-04-24T06:27:14.290000",
      "content": "<p>Try this approach for submission.</p>\n<ul>\n<li>Train your model and save it as an h5 file. Make sure to \"Save output\" while committing your training notebook</li>\n<li>Create a new notebook for submission. Import the output of the training notebook as a dataset, and load the model into your editor</li>\n<li>Use the loaded model to make predictions. </li>\n</ul>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1284221,
      "author_name": "Araik Tamazian",
      "author_url": "",
      "post_date": "2021-04-25T17:42:50.780000",
      "content": "<p>You can speed up computations if you switch to Tensorflow and use TFRecords with augmentations applied beforehand. Check out <a href=\"https://www.kaggle.com/nickuzmenkov/pp2021-tpu-tf-inference\" target=\"_blank\">this notebook</a>.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1282590,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-04-24T05:20:05.227000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1283902,
      "author_name": "Poojan Patel",
      "author_url": "",
      "post_date": "2021-04-25T11:43:38.663000",
      "content": "<p><a href=\"https://www.kaggle.com/mreenav\" target=\"_blank\">@mreenav</a> Thank you for Nice Idea!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1282555": "I think GPU is not used optimally, You can adjust batch size and increase max_queue_size to increase the use of GPU. In my case after adjusting these parameters, one epoch is completed in 10 minutes",
    "1282640": "Try this approach for submission.\n- Train your model and save it as an h5 file. Make sure to \"Save output\" while committing your training notebook\n- Create a new notebook for submission. Import the output of the training notebook as a dataset, and load the model into your editor\n- Use the loaded model to make predictions. ",
    "1284221": "You can speed up computations if you switch to Tensorflow and use TFRecords with augmentations applied beforehand. Check out [this notebook](https://www.kaggle.com/nickuzmenkov/pp2021-tpu-tf-inference).",
    "1282268": "Hello,\nI am using keras, pretrained VGG19 and ImageDataGenerator with flow_from_dataframe method. No extra preprocessing. Still it is taking almost 50 min to complete an epoch. Now notebooks which has GPU time <= 2 hrs are accepted, then how can i submit my notebook?",
    "1282590": "",
    "1283902": "@mreenav Thank you for Nice Idea!"
  }
}