Hi Everyone,
\nI'm having a bit of trouble with the following error when I try to convert my model.
\n\"Some ops are not supported by the native TFLite runtime, you can enable TF kernels fallback using TF Select. See instructions: https://www.tensorflow.org/lite/guide/ops_select
\nTF Select ops: Conv2D, MatrixDiagPartV3\"
The Conv2D is from a convolutional layer I'm using and I'm assuming the \"MatrixDiagPartV3\" is from the following line of code in my notebook: tf.linalg.diag_part(G)
\nWhen I use the following code, which I got from the link provided after \"See instructions:\", the model converts just fine.
\nkeras_model_converter.target_spec.supported_ops = [
\n tf.lite.OpsSet.TFLITE_BUILTINS, # enable TensorFlow Lite ops.
\n tf.lite.OpsSet.SELECT_TF_OPS # enable TensorFlow ops.
\n]
However, as noted in the linked notebook (https://www.kaggle.com/competitions/asl-signs/discussion/398531), these ops don't seem to be supported for this competition…Does anyone have any advice on how to use tf.keras.layers.Conv2D() or tf.linalg.diag_part() in this competition without having to write a convolutional block and a matrix diagonalizer from scratch?
", "messages": [ { "id": "2206734", "postDate": "04/02/2023 19:43:31", "content": "Hi Everyone,
\nI'm having a bit of trouble with the following error when I try to convert my model.
\n\"Some ops are not supported by the native TFLite runtime, you can enable TF kernels fallback using TF Select. See instructions: https://www.tensorflow.org/lite/guide/ops_select
\nTF Select ops: Conv2D, MatrixDiagPartV3\"
The Conv2D is from a convolutional layer I'm using and I'm assuming the \"MatrixDiagPartV3\" is from the following line of code in my notebook: tf.linalg.diag_part(G)
\nWhen I use the following code, which I got from the link provided after \"See instructions:\", the model converts just fine.
\nkeras_model_converter.target_spec.supported_ops = [
\n tf.lite.OpsSet.TFLITE_BUILTINS, # enable TensorFlow Lite ops.
\n tf.lite.OpsSet.SELECT_TF_OPS # enable TensorFlow ops.
\n]
However, as noted in the linked notebook (https://www.kaggle.com/competitions/asl-signs/discussion/398531), these ops don't seem to be supported for this competition…Does anyone have any advice on how to use tf.keras.layers.Conv2D() or tf.linalg.diag_part() in this competition without having to write a convolutional block and a matrix diagonalizer from scratch?
", "rawMarkdown": "Hi Everyone,\n\nI'm having a bit of trouble with the following error when I try to convert my model.\n\n\"Some ops are not supported by the native TFLite runtime, you can enable TF kernels fallback using TF Select. See instructions: https://www.tensorflow.org/lite/guide/ops_select \nTF Select ops: Conv2D, MatrixDiagPartV3\"\n\nThe Conv2D is from a convolutional layer I'm using and I'm assuming the \"MatrixDiagPartV3\" is from the following line of code in my notebook: tf.linalg.diag_part(G)\n\nWhen I use the following code, which I got from the link provided after \"See instructions:\", the model converts just fine.\n\nkeras_model_converter.target_spec.supported_ops = [\n tf.lite.OpsSet.TFLITE_BUILTINS, # enable TensorFlow Lite ops.\n tf.lite.OpsSet.SELECT_TF_OPS # enable TensorFlow ops.\n]\n\nHowever, as noted in the linked notebook (https://www.kaggle.com/competitions/asl-signs/discussion/398531), these ops don't seem to be supported for this competition...Does anyone have any advice on how to use tf.keras.layers.Conv2D() or tf.linalg.diag_part() in this competition without having to write a convolutional block and a matrix diagonalizer from scratch?", "votes": null }, { "id": "2206879", "postDate": "04/03/2023 01:07:17", "content": "https://www.tensorflow.org/model_optimization/guide/pruning/pruning_for_on_device_inference
\nI found this example where tf.keras.layers.Conv2D is acceptable without needing extra OPs, but I can't seem to make it work and when I paste it in my notebook and try to just convert exactly their model, it throws the same error as above…I'm very new to tflite, so this is rather confusing.
", "rawMarkdown": "https://www.tensorflow.org/model_optimization/guide/pruning/pruning_for_on_device_inference\n\nI found this example where tf.keras.layers.Conv2D is acceptable without needing extra OPs, but I can't seem to make it work and when I paste it in my notebook and try to just convert exactly their model, it throws the same error as above...I'm very new to tflite, so this is rather confusing.", "votes": null }, { "id": "2207123", "postDate": "04/03/2023 06:05:28", "content": "https://stackoverflow.com/questions/33700049/get-the-diagonal-of-a-matrix-in-tensorflow
\nx_flat = tf.reshape(x, [-1]) # flatten the matrix\nx_diag = tf.gather(x, [0, 3, 6]) # hard code the digonal index\nIndeed, Conv2D op is supported by tflite, but there may be some limitations. Recently, I've faced the same issue, that Conv2D op was not supported. I've spent a lot of time debugging it and finally I discovered, that key problem was not in Conv2D itself, but in some operation before Conv2D. I've preprocessed data in such a way that I lose the shape of the input feature. This caused a failure of tflite.
\nI am not sure whether this is your case, but I am pretty sure that Conv2D is working well, except for some minor things (like unknown shape, unsupported input data type, etc.). I would recommend you to try to convert a raw model without preprocessing step and check out whether it causes crash.
", "rawMarkdown": "Indeed, Conv2D op is supported by tflite, but there may be some limitations. Recently, I've faced the same issue, that Conv2D op was not supported. I've spent a lot of time debugging it and finally I discovered, that key problem was not in Conv2D itself, but in some operation before Conv2D. I've preprocessed data in such a way that I lose the shape of the input feature. This caused a failure of tflite.\n\nI am not sure whether this is your case, but I am pretty sure that Conv2D is working well, except for some minor things (like unknown shape, unsupported input data type, etc.). I would recommend you to try to convert a raw model without preprocessing step and check out whether it causes crash.", "votes": null }, { "id": "2207475", "postDate": "04/03/2023 12:56:23", "content": "Thanks @Mykola! This might be my problem as well. I'll try debugging this!
", "rawMarkdown": "Thanks @Mykola! This might be my problem as well. I'll try debugging this!", "votes": null }, { "id": "2207667", "postDate": "04/03/2023 15:09:42", "content": "Good luck for you. Would love to hear that was a problem, once you found it
", "rawMarkdown": "Good luck for you. Would love to hear that was a problem, once you found it", "votes": null } ], "comments": [ { "id": 2206879, "author_name": "chemdatafarmer", "author_url": "", "post_date": "04/03/2023 01:07:17", "content": "https://www.tensorflow.org/model_optimization/guide/pruning/pruning_for_on_device_inference
\nI found this example where tf.keras.layers.Conv2D is acceptable without needing extra OPs, but I can't seem to make it work and when I paste it in my notebook and try to just convert exactly their model, it throws the same error as above…I'm very new to tflite, so this is rather confusing.
", "votes": null, "replies": [ { "id": 2207343, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "04/03/2023 10:52:22", "content": "Indeed, Conv2D op is supported by tflite, but there may be some limitations. Recently, I've faced the same issue, that Conv2D op was not supported. I've spent a lot of time debugging it and finally I discovered, that key problem was not in Conv2D itself, but in some operation before Conv2D. I've preprocessed data in such a way that I lose the shape of the input feature. This caused a failure of tflite.
\nI am not sure whether this is your case, but I am pretty sure that Conv2D is working well, except for some minor things (like unknown shape, unsupported input data type, etc.). I would recommend you to try to convert a raw model without preprocessing step and check out whether it causes crash.
", "votes": null, "replies": [ { "id": 2207475, "author_name": "chemdatafarmer", "author_url": "", "post_date": "04/03/2023 12:56:23", "content": "Thanks @Mykola! This might be my problem as well. I'll try debugging this!
", "votes": null, "replies": [ { "id": 2207667, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "04/03/2023 15:09:42", "content": "Good luck for you. Would love to hear that was a problem, once you found it
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2207123, "author_name": "hengck23", "author_url": "", "post_date": "04/03/2023 06:05:28", "content": "https://stackoverflow.com/questions/33700049/get-the-diagonal-of-a-matrix-in-tensorflow
\nx_flat = tf.reshape(x, [-1]) # flatten the matrix\nx_diag = tf.gather(x, [0, 3, 6]) # hard code the digonal index\n