{
  "id": 169154,
  "title": "Is it possible to do transfer learning with DELF model on tfhub?",
  "url": "/competitions/landmark-retrieval-2020/discussion/169154",
  "author_name": "Hirendra Rathor",
  "post_date": "2020-07-23T04:13:34.099000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>The DELF model available on TensorFlow Hub [1] can be used readily for making predictions. However, if I understand correctly, to be able to use this model for transfer learning, I would need to remove the output layers and retrain the model with the training dataset provided in this contest.</p>\n\n<p>Loading this model as a Keras layer does not help since all the layers (including output layers) of the DELF model are available as a single Keras layer.</p>\n\n<p>As mentioned in [2], it appears that if a \"headless\" version (without output layers) of DELF model were available, then the transfer learning could be done with relative ease. I could not find such a version on TF Hub.</p>\n\n<p>Being new to TensorFlow, I am wondering if it is at all possible to use the pretrained model available at [1] to do transfer learning. If it is possible, then any pointers to available documentation will be very much appreciated.</p>\n\n<p><a href=\"/andrefaraujo\">@andrefaraujo</a> could I request your comment please? (assuming you are maintaining the TF Hub package).</p>\n\n<p>Thanks in advance!</p>\n\n<p>[1] <a href=\"https://tfhub.dev/google/delf/1\">https://tfhub.dev/google/delf/1</a>\n[2] <a href=\"https://www.tensorflow.org/tutorials/images/transfer_learning_with_hub#download_the_headless_model\">https://www.tensorflow.org/tutorials/images/transfer_learning_with_hub#download_the_headless_model</a></p>",
  "messages": [
    {
      "id": 940642,
      "postDate": "2020-07-23T04:13:34.100Z",
      "content": "<p>The DELF model available on TensorFlow Hub [1] can be used readily for making predictions. However, if I understand correctly, to be able to use this model for transfer learning, I would need to remove the output layers and retrain the model with the training dataset provided in this contest.</p>\n\n<p>Loading this model as a Keras layer does not help since all the layers (including output layers) of the DELF model are available as a single Keras layer.</p>\n\n<p>As mentioned in [2], it appears that if a \"headless\" version (without output layers) of DELF model were available, then the transfer learning could be done with relative ease. I could not find such a version on TF Hub.</p>\n\n<p>Being new to TensorFlow, I am wondering if it is at all possible to use the pretrained model available at [1] to do transfer learning. If it is possible, then any pointers to available documentation will be very much appreciated.</p>\n\n<p><a href=\"/andrefaraujo\">@andrefaraujo</a> could I request your comment please? (assuming you are maintaining the TF Hub package).</p>\n\n<p>Thanks in advance!</p>\n\n<p>[1] <a href=\"https://tfhub.dev/google/delf/1\">https://tfhub.dev/google/delf/1</a>\n[2] <a href=\"https://www.tensorflow.org/tutorials/images/transfer_learning_with_hub#download_the_headless_model\">https://www.tensorflow.org/tutorials/images/transfer_learning_with_hub#download_the_headless_model</a></p>",
      "rawMarkdown": "The DELF model available on TensorFlow Hub [1] can be used readily for making predictions. However, if I understand correctly, to be able to use this model for transfer learning, I would need to remove the output layers and retrain the model with the training dataset provided in this contest.\n\nLoading this model as a Keras layer does not help since all the layers (including output layers) of the DELF model are available as a single Keras layer.\n\nAs mentioned in [2], it appears that if a \"headless\" version (without output layers) of DELF model were available, then the transfer learning could be done with relative ease. I could not find such a version on TF Hub.\n\nBeing new to TensorFlow, I am wondering if it is at all possible to use the pretrained model available at [1] to do transfer learning. If it is possible, then any pointers to available documentation will be very much appreciated.\n\n@andrefaraujo could I request your comment please? (assuming you are maintaining the TF Hub package).\n\nThanks in advance!\n\n[1] https://tfhub.dev/google/delf/1\n[2] https://www.tensorflow.org/tutorials/images/transfer_learning_with_hub#download_the_headless_model"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "940642": "The DELF model available on TensorFlow Hub [1] can be used readily for making predictions. However, if I understand correctly, to be able to use this model for transfer learning, I would need to remove the output layers and retrain the model with the training dataset provided in this contest.\n\nLoading this model as a Keras layer does not help since all the layers (including output layers) of the DELF model are available as a single Keras layer.\n\nAs mentioned in [2], it appears that if a \"headless\" version (without output layers) of DELF model were available, then the transfer learning could be done with relative ease. I could not find such a version on TF Hub.\n\nBeing new to TensorFlow, I am wondering if it is at all possible to use the pretrained model available at [1] to do transfer learning. If it is possible, then any pointers to available documentation will be very much appreciated.\n\n@andrefaraujo could I request your comment please? (assuming you are maintaining the TF Hub package).\n\nThanks in advance!\n\n[1] https://tfhub.dev/google/delf/1\n[2] https://www.tensorflow.org/tutorials/images/transfer_learning_with_hub#download_the_headless_model"
  }
}