{
  "id": 280584,
  "title": "Can't downgrade from Tensorflow 2.6.0 to Tensorflow 2.2.0",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/280584",
  "author_name": "",
  "post_date": "2021-10-21T20:01:50.279962500Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I made a model in Tensorflow 2.2.0 (with Keras 2.3.1) in Colab but failed to do the same in Kaggle.<br>\nI downgraded TF and installed TF-2.2.0 and Keras-2.3.1 in Kaggle but doing this it does not register the GPU…it runs only on the CPU and that is making an epoch train time from 30 min to 8 hours, <br>\nIt returns this message before starting the training…<br>\nrefer to this notebook: <a href=\"https://www.kaggle.com/susnato/mask-rcnn-tf-train\" target=\"_blank\">https://www.kaggle.com/susnato/mask-rcnn-tf-train</a></p>\n<blockquote>\n  <p>Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at <a href=\"https://www.tensorflow.org/install/gpu\" target=\"_blank\">https://www.tensorflow.org/install/gpu</a></p>\n</blockquote>",
  "messages": [
    {
      "id": "1553020",
      "postDate": "10/21/2021 20:01:50",
      "content": "<p>I made a model in Tensorflow 2.2.0 (with Keras 2.3.1) in Colab but failed to do the same in Kaggle.<br>\nI downgraded TF and installed TF-2.2.0 and Keras-2.3.1 in Kaggle but doing this it does not register the GPU…it runs only on the CPU and that is making an epoch train time from 30 min to 8 hours, <br>\nIt returns this message before starting the training…<br>\nrefer to this notebook: <a href=\"https://www.kaggle.com/susnato/mask-rcnn-tf-train\" target=\"_blank\">https://www.kaggle.com/susnato/mask-rcnn-tf-train</a></p>\n<blockquote>\n  <p>Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at <a href=\"https://www.tensorflow.org/install/gpu\" target=\"_blank\">https://www.tensorflow.org/install/gpu</a></p>\n</blockquote>",
      "rawMarkdown": "I made a model in Tensorflow 2.2.0 (with Keras 2.3.1) in Colab but failed to do the same in Kaggle.\nI downgraded TF and installed TF-2.2.0 and Keras-2.3.1 in Kaggle but doing this it does not register the GPU...it runs only on the CPU and that is making an epoch train time from 30 min to 8 hours, \nIt returns this message before starting the training...\nrefer to this notebook: https://www.kaggle.com/susnato/mask-rcnn-tf-train\n\n> Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu",
      "votes": null
    },
    {
      "id": "1565030",
      "postDate": "10/30/2021 00:35:29",
      "content": "<p>Getting tensorflow to work on GPU requires the correct cudnn and cudatoolkit versions - not sure if I have seen any shared kernels on any competitions that show install of these two.  </p>\n<p>The correct versions can be found buried in the tensorflow web site - can't seem to find that page right now.  </p>\n<p>I run 4 local machines and due a full fresh install of Ubuntu and Anaconda for each competition and upfront I need to decide what tf version (like to do same as kaggle) so I can install these two.  I use this bit of code after install to confirm that the gpus are being seen.</p>\n<p>import tensorflow as tf<br>\nprint(\"Num GPUs Available: \", len(tf.config.list_physical_devices('GPU')))<br>\ngpus = tf.config.list_physical_devices('GPU')<br>\ngpus</p>",
      "rawMarkdown": "Getting tensorflow to work on GPU requires the correct cudnn and cudatoolkit versions - not sure if I have seen any shared kernels on any competitions that show install of these two.  \n\nThe correct versions can be found buried in the tensorflow web site - can't seem to find that page right now.  \n\nI run 4 local machines and due a full fresh install of Ubuntu and Anaconda for each competition and upfront I need to decide what tf version (like to do same as kaggle) so I can install these two.  I use this bit of code after install to confirm that the gpus are being seen.\n\nimport tensorflow as tf\nprint(\"Num GPUs Available: \", len(tf.config.list_physical_devices('GPU')))\ngpus = tf.config.list_physical_devices('GPU')\ngpus",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1565030,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "10/30/2021 00:35:29",
      "content": "<p>Getting tensorflow to work on GPU requires the correct cudnn and cudatoolkit versions - not sure if I have seen any shared kernels on any competitions that show install of these two.  </p>\n<p>The correct versions can be found buried in the tensorflow web site - can't seem to find that page right now.  </p>\n<p>I run 4 local machines and due a full fresh install of Ubuntu and Anaconda for each competition and upfront I need to decide what tf version (like to do same as kaggle) so I can install these two.  I use this bit of code after install to confirm that the gpus are being seen.</p>\n<p>import tensorflow as tf<br>\nprint(\"Num GPUs Available: \", len(tf.config.list_physical_devices('GPU')))<br>\ngpus = tf.config.list_physical_devices('GPU')<br>\ngpus</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1553020": "I made a model in Tensorflow 2.2.0 (with Keras 2.3.1) in Colab but failed to do the same in Kaggle.\nI downgraded TF and installed TF-2.2.0 and Keras-2.3.1 in Kaggle but doing this it does not register the GPU...it runs only on the CPU and that is making an epoch train time from 30 min to 8 hours, \nIt returns this message before starting the training...\nrefer to this notebook: https://www.kaggle.com/susnato/mask-rcnn-tf-train\n\n> Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu",
    "1565030": "Getting tensorflow to work on GPU requires the correct cudnn and cudatoolkit versions - not sure if I have seen any shared kernels on any competitions that show install of these two.  \n\nThe correct versions can be found buried in the tensorflow web site - can't seem to find that page right now.  \n\nI run 4 local machines and due a full fresh install of Ubuntu and Anaconda for each competition and upfront I need to decide what tf version (like to do same as kaggle) so I can install these two.  I use this bit of code after install to confirm that the gpus are being seen.\n\nimport tensorflow as tf\nprint(\"Num GPUs Available: \", len(tf.config.list_physical_devices('GPU')))\ngpus = tf.config.list_physical_devices('GPU')\ngpus"
  },
  "source": "meta"
}