Hi, If anyone's not working on this, then I would like to work on it.

Hi, is it worth waiting for implementation FakeQuantWithMinMaxVarsPerChannel?

Hi all,

As we work toward fleshing out the builtin op library for TensorFlow Lite, we've been working on an experimental feature that allows using select TensorFlow ops from within the TensorFlow Lite runtime. The goal is to help reduce some of the friction for using models that rely on ops not yet natively supported by TensorFlow Lite (at the cost of increased binary size). This feature requires opting in during model conversion, as well as adding an additional dependency. More details can be found here.

Feedback is very much appreciated (either via GitHub or directly via tflite@tensorflow.org), and we'll be adding and refining functionality over the coming weeks. Cheers.

Hi, is it worth waiting for implementation FakeQuantWithMinMaxVarsPerChannel?

It is unlikely that we'll adding additional support for FakeQuant ops in the near future. Your best bet is to look into using post-training quantization.

@suharshs @jdduke Hi, I want to know when conv3d feature will be release?

Hi @StephenLee2016 Conv3D is one of the select TF ops supported via tflite_convert.

Hi all,

As we work toward fleshing out the builtin op library for TensorFlow Lite, we've been working on an experimental feature that allows using select TensorFlow ops from within the TensorFlow Lite runtime. The goal is to help reduce some of the friction for using models that rely on ops not yet natively supported by TensorFlow Lite (at the cost of increased binary size). This feature requires opting in during model conversion, as well as adding an additional dependency. More details can be found here.

Feedback is very much appreciated (either via GitHub or directly via tflite@tensorflow.org), and we'll be adding and refining functionality over the coming weeks. Cheers.

Hi all,

As we work toward fleshing out the builtin op library for TensorFlow Lite, we've been working on an experimental feature that allows using select TensorFlow ops from within the TensorFlow Lite runtime. The goal is to help reduce some of the friction for using models that rely on ops not yet natively supported by TensorFlow Lite (at the cost of increased binary size). This feature requires opting in during model conversion, as well as adding an additional dependency. More details can be found here.

Feedback is very much appreciated (either via GitHub or directly via tflite@tensorflow.org), and we'll be adding and refining functionality over the coming weeks. Cheers.

TF version is 1.12.0， I can not find target_ops options in tflite_convert, why?

@ToBigboss Please note that you will need to build from source to gain early access to the new features.

@Lucy20211

Is it worth waiting for implementation of Conv1D? I know that I can use Conv2D instead of Conv1D, but I have to measure performance of my MCU which uses the same Conv1D and make a comparison with Conv2D for different models (if I implemented Conv1D with Conv2D, it should use the same resources and complexity of a Conv2D, so I wouldn't do anything).

Facing the same issue, only that I don't know how to use Conv2D instead of Conv1D. Could you point me to some tutorial/reference?

Hi @mirkomartn,

what you could do is well summarized in the last comment here: #43141.

@Lucy20211 Great, thank you!

Wanted to add tf.Selu to the list of tflite ops for promotion. It is available using Select TF Ops but will be nice to have it as a built-in since we already have ELU. Thanks!

#50595 requests BroadcastGradientArgs, DynamicStitch, EluGrad, Sign, StridedSliceGrad, UnsortedSegmentSum

Can tf.einsum be supported?

Can tf.einsum be supported?

tf.einsum with static shape is fully supported via our converter. I think you need dynamic shape, right?

Hi,
Could you please assist with AssignVariableOp, ReadVariableOp, VarHandleOp ? Is it worth waiting?
I know it's possible to use it with SELECT_TF_OPS, but such an option increases dll size significantly, so it would be nice to have it as a build-in. Thanks!

Failed porting TensorListFromTensor as a custom op. Is it going to be added? I know I can use it through select ops, which will create a Flex version with the new experimental converter. But the .aar file created for Android gets too large

Would be useful to have:

AvgPool3D, MaxPool3D, TensorListFromTensor, TensorListGetItem, TensorListReserve, TensorListSetItem, TensorListStack

Model is I3D with a custom top layer.

Hi @suharshs !
I think Quantized Div error has been taken care of in the 2.8 version now. Attached gist and relevant thread 1 and 2 for reference.
Thank you!

I want to let my model .pb file to .tflite, and run on edge device.
But I have an error about the "MatrixBandPart" Ops, I want to ask this Ops is still not support?

The error is :
:0: error: failed while converting: 'main': Ops that need custom implementation (enabled via setting the -emit-custom-ops flag):
tf.MatrixBandPart {device = ""}

I also get some trouble about :
WARNING:absl:Found untraced functions such as dense_32_layer_call_and_return_conditional_losses, dense_32_layer_call_fn, embedding_layer_call_and_return_conditional_losses, embedding_layer_call_fn, dropout_layer_call_and_return_conditional_losses while saving (showing 5 of 295). These functions will not be directly callable after loading.

Do you have any solution of these error?

@lp0617 !
The Above warning does not impact much TFLiteconversion. To use custom op, You need to register the op in TFlite kernel and use custom ops syntax during TFLiteconversion.
converter.allow_custom_ops = True

Attached relevant thread for reference.

Thank you!

Since all the issues mentioned in the task is closed, cloud you please close this issue. Thanks!

This issue has been automatically marked as stale because it has no recent activity. It will be closed if no further activity occurs. Thank you.

Closing as stale. Please reopen if you'd like to work on this further.

tf.raw_ops.MatrixInverse is not supported in tflite, while Op BatchMatrixInverse is not available in GraphDef version 1205. How can i calculate the inverse of the matrix in tflite?

request to add tf.keras.layers.TextVectorization in tflite ops

TensorListReserve would be much appreciated

Request to add tf.RealDiv and tf.erf please. Thanks!

Request to add Recurrent Layers (GRU, LSTM).