{
  "id": 208253,
  "title": "Keeping the tensorflow consistent",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/208253",
  "author_name": "Mahdi Hosseinali",
  "post_date": "2021-01-02T15:07:31.540000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>When you activate the TPUs, the TensorFlow version is 2.2.0 but for the GPU/CPU accelerators, the TensorFlow version is 2.3.1. The tf version in the TPU is limited by google as they should be installed on the TPUs and probably for the normal accelerator the docker is the most up to date. But this means that I cannot switch between accelerators and transfer my weights. Is there any decent way to downgrade the tf 2.3.1 to 2.2.0?</p>\n<p>PS: restarting the session without the kernel as suggested <a href=\"https://www.kaggle.com/questions-and-answers/173815\" target=\"_blank\">here</a> is not a solution because it only works in interactive sessions. I tried to uninstall and then install but doesn't work either, the installation fails because of tensorflow-cloud and when I check the version it is 2.3.1</p>",
  "messages": [
    {
      "id": 1135872,
      "postDate": "2021-01-02T15:07:31.540Z",
      "content": "<p>When you activate the TPUs, the TensorFlow version is 2.2.0 but for the GPU/CPU accelerators, the TensorFlow version is 2.3.1. The tf version in the TPU is limited by google as they should be installed on the TPUs and probably for the normal accelerator the docker is the most up to date. But this means that I cannot switch between accelerators and transfer my weights. Is there any decent way to downgrade the tf 2.3.1 to 2.2.0?</p>\n<p>PS: restarting the session without the kernel as suggested <a href=\"https://www.kaggle.com/questions-and-answers/173815\" target=\"_blank\">here</a> is not a solution because it only works in interactive sessions. I tried to uninstall and then install but doesn't work either, the installation fails because of tensorflow-cloud and when I check the version it is 2.3.1</p>",
      "rawMarkdown": "When you activate the TPUs, the TensorFlow version is 2.2.0 but for the GPU/CPU accelerators, the TensorFlow version is 2.3.1. The tf version in the TPU is limited by google as they should be installed on the TPUs and probably for the normal accelerator the docker is the most up to date. But this means that I cannot switch between accelerators and transfer my weights. Is there any decent way to downgrade the tf 2.3.1 to 2.2.0?\n\nPS: restarting the session without the kernel as suggested [here](https://www.kaggle.com/questions-and-answers/173815) is not a solution because it only works in interactive sessions. I tried to uninstall and then install but doesn't work either, the installation fails because of tensorflow-cloud and when I check the version it is 2.3.1"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1135872": "When you activate the TPUs, the TensorFlow version is 2.2.0 but for the GPU/CPU accelerators, the TensorFlow version is 2.3.1. The tf version in the TPU is limited by google as they should be installed on the TPUs and probably for the normal accelerator the docker is the most up to date. But this means that I cannot switch between accelerators and transfer my weights. Is there any decent way to downgrade the tf 2.3.1 to 2.2.0?\n\nPS: restarting the session without the kernel as suggested [here](https://www.kaggle.com/questions-and-answers/173815) is not a solution because it only works in interactive sessions. I tried to uninstall and then install but doesn't work either, the installation fails because of tensorflow-cloud and when I check the version it is 2.3.1"
  }
}