The submission environment does not support TensorFlow operators.
\nIt seems to have been the case in the previous competition as mentioned here.
\nThis means for a submission to work, one should be able to convert their model to TFLite without using the tf.lite.OpsSet.SELECT_TF_OPS flag.
keras_model_converter = tf.lite.TFLiteConverter.from_keras_model(model)\n\nkeras_model_converter.target_spec.supported_ops = [\n tf.lite.OpsSet.TFLITE_BUILTINS, \n tf.lite.OpsSet.SELECT_TF_OPS \n]\ntflite_model = keras_model_converter.convert()\nThe problem is that without the SELECT_TF_OPSflag, we are very limited in building our model.
We can't use simple layers such as tf.keras.layers.Embedding, and simple operations like tf.gather or tf.gather_nd are not working, which makes it impossible to perform indexing on tensors.
There is already constraints to guarantee that submitted models could potentially run on cell phones:
\nHence, we kindly request the organizers to include TensorFlow operators in the submission environment and allow contestants to ensure that their models satisfy the runtime constraints.
\nPlease upvote this thread if you agree with me😓.
", "messages": [ { "id": 2275054, "postDate": "2023-05-26T13:06:00.133Z", "content": "The submission environment does not support TensorFlow operators.
\nIt seems to have been the case in the previous competition as mentioned here.
\nThis means for a submission to work, one should be able to convert their model to TFLite without using the tf.lite.OpsSet.SELECT_TF_OPS flag.
keras_model_converter = tf.lite.TFLiteConverter.from_keras_model(model)\n\nkeras_model_converter.target_spec.supported_ops = [\n tf.lite.OpsSet.TFLITE_BUILTINS, \n tf.lite.OpsSet.SELECT_TF_OPS \n]\ntflite_model = keras_model_converter.convert()\nThe problem is that without the SELECT_TF_OPSflag, we are very limited in building our model.
We can't use simple layers such as tf.keras.layers.Embedding, and simple operations like tf.gather or tf.gather_nd are not working, which makes it impossible to perform indexing on tensors.
There is already constraints to guarantee that submitted models could potentially run on cell phones:
\nHence, we kindly request the organizers to include TensorFlow operators in the submission environment and allow contestants to ensure that their models satisfy the runtime constraints.
\nPlease upvote this thread if you agree with me😓.
", "rawMarkdown": "The submission environment does not support TensorFlow operators.\n\nIt seems to have been the case in the previous competition as mentioned [here](https://www.kaggle.com/competitions/asl-signs/discussion/390008#2199328).\n\nThis means for a submission to work, one should be able to convert their model to TFLite without using the `tf.lite.OpsSet.SELECT_TF_OPS` flag.\n\n\n```python\nkeras_model_converter = tf.lite.TFLiteConverter.from_keras_model(model)\n\nkeras_model_converter.target_spec.supported_ops = [\n tf.lite.OpsSet.TFLITE_BUILTINS, \n tf.lite.OpsSet.SELECT_TF_OPS # You can't use this\n]\ntflite_model = keras_model_converter.convert()\n\n```\n\nThe problem is that without the `SELECT_TF_OPS `flag, we are very limited in building our model.\n\nWe can't use simple layers such as `tf.keras.layers.Embedding`, and simple operations like `tf.gather` or `tf.gather_nd` are not working, which makes it impossible to perform indexing on tensors.\n\nThere is already constraints to guarantee that submitted models could potentially run on cell phones:\n\n- Your model must perform inference in less than 5 hours.\n- Your model must use less than 40 MB of storage space.\n\nHence, we kindly request the organizers to include **TensorFlow operators** in the submission environment and allow contestants to ensure that their models satisfy the runtime constraints.\n\nPlease upvote this thread if you agree with me😓.", "votes": 10 }, { "id": 2276212, "postDate": "2023-05-26T15:21:26.983Z", "content": "Meanwhile, I can confirm that tf.gather works well in TFLIte without TF_OPS (verified in the previous competition)
Yeah I know but it has many weird bugs that can appear without any explanation. Check this GitHub issue: https://github.com/tensorflow/tensorflow/issues/42410
", "rawMarkdown": "Yeah I know but it has many weird bugs that can appear without any explanation. Check this GitHub issue: https://github.com/tensorflow/tensorflow/issues/42410" } ] }, { "id": 2276193, "postDate": "2023-05-26T15:12:16.843Z", "content": "It is also important to note that you may include tf.lite.OpsSet.SELECT_TF_OPS but if not used in the model the submission will be successful as in this notebook: https://www.kaggle.com/code/chack3/pytorch-simple-model
", "rawMarkdown": "It is also important to note that you may include tf.lite.OpsSet.SELECT_TF_OPS but if not used in the model the submission will be successful as in this notebook: https://www.kaggle.com/code/chack3/pytorch-simple-model", "votes": 1 }, { "id": 2276233, "postDate": "2023-05-26T15:36:02.563Z", "content": "Thanks for sharing the convert code, it can helped to training the model in different form.
", "rawMarkdown": "Thanks for sharing the convert code, it can helped to training the model in different form." }, { "id": 2301727, "postDate": "2023-06-14T06:13:28.133Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2276230, "postDate": "2023-05-26T15:34:10.573Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2275101, "postDate": "2023-05-26T14:01:38.600Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2276212, "author_name": "Mykola", "author_url": "", "post_date": "2023-05-26T15:21:26.983000", "content": "Meanwhile, I can confirm that tf.gather works well in TFLIte without TF_OPS (verified in the previous competition)
Yeah I know but it has many weird bugs that can appear without any explanation. Check this GitHub issue: https://github.com/tensorflow/tensorflow/issues/42410
", "votes": 0, "replies": [] } ] }, { "id": 2276193, "author_name": "Jude TCHAYE", "author_url": "", "post_date": "2023-05-26T15:12:16.843000", "content": "It is also important to note that you may include tf.lite.OpsSet.SELECT_TF_OPS but if not used in the model the submission will be successful as in this notebook: https://www.kaggle.com/code/chack3/pytorch-simple-model
", "votes": 1, "replies": [] }, { "id": 2276233, "author_name": "Dewei Chen", "author_url": "", "post_date": "2023-05-26T15:36:02.563000", "content": "Thanks for sharing the convert code, it can helped to training the model in different form.
", "votes": 0, "replies": [] }, { "id": 2301727, "author_name": "", "author_url": "", "post_date": "2023-06-14T06:13:28.133000", "content": "", "votes": 0, "replies": [] }, { "id": 2276230, "author_name": "", "author_url": "", "post_date": "2023-05-26T15:34:10.573000", "content": "", "votes": 0, "replies": [] }, { "id": 2275101, "author_name": "", "author_url": "", "post_date": "2023-05-26T14:01:38.600000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2275054": "The submission environment does not support TensorFlow operators.\n\nIt seems to have been the case in the previous competition as mentioned [here](https://www.kaggle.com/competitions/asl-signs/discussion/390008#2199328).\n\nThis means for a submission to work, one should be able to convert their model to TFLite without using the `tf.lite.OpsSet.SELECT_TF_OPS` flag.\n\n\n```python\nkeras_model_converter = tf.lite.TFLiteConverter.from_keras_model(model)\n\nkeras_model_converter.target_spec.supported_ops = [\n tf.lite.OpsSet.TFLITE_BUILTINS, \n tf.lite.OpsSet.SELECT_TF_OPS # You can't use this\n]\ntflite_model = keras_model_converter.convert()\n\n```\n\nThe problem is that without the `SELECT_TF_OPS `flag, we are very limited in building our model.\n\nWe can't use simple layers such as `tf.keras.layers.Embedding`, and simple operations like `tf.gather` or `tf.gather_nd` are not working, which makes it impossible to perform indexing on tensors.\n\nThere is already constraints to guarantee that submitted models could potentially run on cell phones:\n\n- Your model must perform inference in less than 5 hours.\n- Your model must use less than 40 MB of storage space.\n\nHence, we kindly request the organizers to include **TensorFlow operators** in the submission environment and allow contestants to ensure that their models satisfy the runtime constraints.\n\nPlease upvote this thread if you agree with me😓.", "2276212": "Meanwhile, I can confirm that `tf.gather` works well in TFLIte without TF_OPS (verified in the previous competition)", "2276193": "It is also important to note that you may include tf.lite.OpsSet.SELECT_TF_OPS but if not used in the model the submission will be successful as in this notebook: https://www.kaggle.com/code/chack3/pytorch-simple-model", "2276233": "Thanks for sharing the convert code, it can helped to training the model in different form.", "2301727": "", "2276230": "", "2275101": "" } }