Run ONNX models on your Android phone using the new NNAPI !
Android 8.1 introduces Neural Networks API (NNAPI). It's very exciting to run a model in the "native" way supported by Android System. :)
DNNLibrary is a wrapper of NNAPI ("DNNLibrary" is for "daquexian's NNAPI library). It lets you easily make the use of the new NNAPI introduced in Android 8.1. You can convert your onnx model into
daq and run the model directly.
For the Android app example, please check out dnnlibrary-example.
Telegram Group: link, QQ Group (Chinese): 948989771, answer: 哈哈哈哈
This screenshot is MobileNet v2, both float version and 8-bit quantized version
Clone this repo and submodules:
git clone --recursive https://github.com/daquexian/DNNLibrary
Please make sure the Android System on your phone is 8.1+, or you may want to use an 8.1+ emulator.
Android 8.1 introduces NNAPI. However, NNAPI is not friendly to normal Android developers. It is not designed to be used by normal developers directly. So I wrapped it into a library.
With DNNLibrary it's extremely easy to deploy your ONNX model on Android 8.1+ phone. For example, following is the Java code to deploy the MobileNet v2 in your app (please check out dnnlibrary-example for detail):
ModelBuilder modelBuilder = new ModelBuilder(); Model model = modelBuilder.readFile(getAssets(), "mobilenetv2.daq") // the following line will allow fp16 on supported devices, bringing speed boost. It is only available on Android P, see https://www.anandtech.com/show/13503/the-mate-20-mate-20-pro-review/4 for a detailed benchmark // .allowFp16(true) .setOutput("mobilenetv20_output_pred_fwd"); // The output name is from the onnx model .compile(ModelBuilder.PREFERENCE_FAST_SINGLE_ANSWER); float result = model.predict(inputData);
Only five lines! And the
daq model file is got from the pretrained onnx model using
Convert the model
After cloning step listed in Preparation section, run
mkdir build cd build cmake .. cmake --build .
onnx2daq is in
tools/onnx2daq directory. The following command is to convert onnx model to daq model.
./tools/onnx2daq/onnx2daq onnx_model output_filename
For example, if you have a model named "mobilenetv2.onnx" in your current directory,
./tools/onnx2daq/onnx2daq mobilenetv2.onnx mobilenetv2.daq
For 8-bit quantization, please check out our wiki
If you are an Android app developer and want it to work out of the box
Welcome! It has been published on jcenter.
in your app's
The lastest version can be found in the following badge:
If you are an c++ developer and don't care about Android app
We use CMake as the build system. So you can build it as most C++ projects, the only difference is that you need Android NDK, r17b or higher NDK is necessary :
mkdir build && cd build cmake -DCMAKE_SYSTEM_NAME=Android -DCMAKE_TOOLCHAIN_FILE=path_of_android_ndk/build/cmake/android.toolchain.cmake -DANDROID_CPP_FEATURES=exceptions -DANDROID_PLATFORM=replace_me_with_android-28_or_android-27 -DANDROID_ABI=arm64-v8a cmake --build .
then you will get binary files.
But TensorFlow Lite also supports NNAPI...
Yes, but its support for NNAPI is far from perfect. For example, dilated convolution (which is widely used in segmentation) are not supported, prelu is also not supported.
What's more, only the TensorFlow models can easily get converted to TensorFlow Lite model. Since NNAPI is independent of any frameworks, we support ONNX, a framework-independent model format.
|Supported Model Format||TensorFlow||ONNX|
|Ease of Use||❌
(Bazel build system,
not friendly to Android developers)
However we are also far from maturity comparing to TF Lite. At least we are an another choice if you want to enjoy the power of NNAPI :)
|Google Pixel, API 28||MobileNet v2||132.95ms|
|Google Pixel, API 28||SqueezeNet v1.1||80.80ms|
|RK3399, API 27||MobileNet v2||59.90ms|
|RK3399, API 27||SqueezeNet v1.1||51.10ms|
More benchmark is welcome!
About caffe model support
The old DNNLibrary supports caffe model by dnntools, however, it is not supported directly now, the models generated by
dnntools are not usable, too. Please use a convert tool like MMdnn to convert the caffe model to the ONNX model, then convert it to