Caution: This feature is experimental.
The TensorFlow Lite builtin op library has grown rapidly, and will continue to grow, but there remains a long tail of TensorFlow ops that are not yet natively supported by TensorFlow Lite . These unsupported ops can be a point of friction in the TensorFlow Lite model conversion process. To that end, the team has recently been working on an experimental mechanism for reducing this friction.
This document outlines how to use TensorFlow Lite with select TensorFlow ops. Note that this feature is experimental and is under active development. As you use this feature, keep in mind the known limitations, and please send feedback about models that work and issues you are facing to tflite@tensorflow.org.
TensorFlow Lite will continue to have TensorFlow Lite builtin ops optimized for mobile and embedded devices. However, TensorFlow Lite models can now use a subset of TensorFlow ops when TFLite builtin ops are not sufficient.
Models converted with TensorFlow ops will require a TensorFlow Lite interpreter that has a larger binary size than the interpreter with only TFLite builtin ops. Additionally, performance optimizations will not be available for any TensorFlow ops in the TensorFlow Lite model.
This document outlines how to convert and run a TFLite model with TensorFlow ops on your platform of choice. It also discusses some known limitations, the future plans for this feature, and basic performance and size metrics.
To convert a TensorFlow model to a TensorFlow Lite model with TensorFlow ops, use the target_spec.supported_ops argument in the TensorFlow Lite converter. The following values are valid options for target_spec.supported_ops:
TFLITE_BUILTINS - Converts models using TensorFlow Lite builtin ops.SELECT_TF_OPS - Converts models using TensorFlow ops. The exact subset of supported ops can be found in the whitelist at lite/delegates/flex/whitelisted_flex_ops.cc.Note: target_spec.supported_ops was previously target_ops in the Python API.
The recommended approach is to convert the model with TFLITE_BUILTINS, then with both TFLITE_BUILTINS,SELECT_TF_OPS, and finally with only SELECT_TF_OPS. Using both options (i.e. TFLITE_BUILTINS,SELECT_TF_OPS) creates models with TensorFlow Lite ops where possible. Using only SELECT_TF_OPS is useful when the model contains TensorFlow ops that are only partially supported by TensorFlow Lite, and one would like to avoid those limitations.
The following example shows how to use this feature in the TFLiteConverter Python API.
import tensorflow as tf
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS,
tf.lite.OpsSet.SELECT_TF_OPS]
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)
The following example shows how to use this feature in the tflite_convert command line tool using the command line flag target_ops.
tflite_convert \ --output_file=/tmp/foo.tflite \ --graph_def_file=/tmp/foo.pb \ --input_arrays=input \ --output_arrays=MobilenetV1/Predictions/Reshape_1 \ --target_ops=TFLITE_BUILTINS,SELECT_TF_OPS
When building and running tflite_convert directly with bazel, please pass --define=with_select_tf_ops=true as an additional argument.
bazel run --define=with_select_tf_ops=true tflite_convert -- \ --output_file=/tmp/foo.tflite \ --graph_def_file=/tmp/foo.pb \ --input_arrays=input \ --output_arrays=MobilenetV1/Predictions/Reshape_1 \ --target_ops=TFLITE_BUILTINS,SELECT_TF_OPS
When using a TensorFlow Lite model that has been converted with support for select TensorFlow ops, the client must also use a TensorFlow Lite runtime that includes the necessary library of TensorFlow ops.
A new Android AAR target with select TensorFlow ops has been added for convenience. Assuming a working TensorFlow Lite build environment, build the Android AAR with select TensorFlow ops as follows:
bazel build --cxxopt='--std=c++11' -c opt \ --config=android_arm --config=monolithic \ //tensorflow/lite/java:tensorflow-lite-with-select-tf-ops
This will generate an AAR file in bazel-genfiles/tensorflow/lite/java/. From there, you can either import the AAR directly into your project, or publish the custom AAR to your local Maven repository:
mvn install:install-file \ -Dfile=bazel-genfiles/tensorflow/lite/java/tensorflow-lite-with-select-tf-ops.aar \ -DgroupId=org.tensorflow \ -DartifactId=tensorflow-lite-with-select-tf-ops -Dversion=0.1.100 -Dpackaging=aar
Finally, in your app's build.gradle, ensure you have the mavenLocal() dependency and replace the standard TensorFlow Lite dependency with the one that has support for select TensorFlow ops:
allprojects {
repositories {
jcenter()
mavenLocal()
}
}
dependencies {
implementation 'org.tensorflow:tensorflow-lite-with-select-tf-ops:0.1.100'
}
With XCode Command Line Tools installed, TensorFlow Lite with select TensorFlow ops support can be built with the following command:
tensorflow/contrib/makefile/build_all_ios_with_tflite.sh
This will generate the required static linking libraries in the tensorflow/contrib/makefile/gen/lib/ directory.
The TensorFlow Lite camera example app can be used to test this. A new TensorFlow Lite XCode project with support for select TensorFlow ops has been added to tensorflow/lite/examples/ios/camera/tflite_camera_example_with_select_tf_ops.xcodeproj.
To use this feature in your own project, either clone the example project or set the project settings for a new or existing project to the following:
tensorflow/contrib/makefile/gen/lib/ directory:libtensorflow-lite.alibprotobuf.ansync.atensorflow/lite/tensorflow/contrib/makefile/downloads/flatbuffer/includetensorflow/contrib/makefile/downloads/eigen-force_load tensorflow/contrib/makefile/gen/lib/libtensorflow-lite.a.A CocoaPod with support for select TensorFlow ops will also be released in the future.
When building TensorFlow Lite libraries using the bazel pipeline, the additional TensorFlow ops library can be included and enabled as follows:
--config=monolithic build flag.--define=with_select_tf_ops=true build flag in the bazel build invocation when building TensorFlow Lite.tensorflow/lite/delegates/flex:delegate.Note that the necessary TfLiteDelegate will be installed automatically when creating the interpreter at runtime as long as the delegate is linked into the client library. It is not necessary to explicitly install the delegate instance as is typically required with other delegate types.
Python support is actively under development.
When using a mixture of both builtin and select TensorFlow ops, all of the same TensorFlow Lite optimizations and optimized builtin kernels will be be available and usable with the converted model.
The following table describes the average time taken to run inference on MobileNet on a Pixel 2. The listed times are an average of 100 runs. These targets were built for Android using the flags: --config=android_arm64 -c opt.
| Build | Time (milliseconds) |
|---|---|
Only built-in ops (TFLITE_BUILTIN) | 260.7 |
Using only TF ops (SELECT_TF_OPS) | 264.5 |
The following table describes the binary size of TensorFlow Lite for each build. These targets were built for Android using --config=android_arm -c opt.
| Build | C++ Binary Size | Android APK Size |
|---|---|---|
| Only built-in ops | 796 KB | 561 KB |
| Built-in ops + TF ops | 23.0 MB | 8.0 MB |
The following is a list of some of the known limitations:
post_training_quantization flag is currently not supported for TensorFlow ops so it will not quantize weights for any TensorFlow ops. In models with both TensorFlow Lite builtin ops and TensorFlow ops, the weights for the builtin ops will be quantized.The following is a list of improvements to this pipeline that are in progress: