Building TensorFlow Lite Support Library with Docker

Written by George Soloupis ML GDE

This is an extended guide that illustrates how to build Tensorflow Lite Support library with Docker inside Windows. This procedure is useful to developers that want to make changes to the library but they haven’t installed Bazel to their system. After build procedure the developer can obtain an .aar file that can be used inside an Android Studio project.

First install Docker

If you have already Docker in your system then you can skip this part. If not go to the official page, download Docker for Windows and install Docker in your system. Run Docker desktop and if it prompts to upgrade WSL 2 then do this procedure also. After a restart of your PC run again to start Docker. You will see at your desktop the main window:

Collect neccessary files and start a container

Create a new folder and insert the 3 important files (build_support_aar_with_docker.sh, build_support_aar.sh and tflite-android.Dockerfile) that you can find at the main branch of this github repository:

Click Shift + Right click of mouse and open a PowerShell window inside the folder that contains the 3 files:

Then copy, paste command:

docker build . -t tf-support-builder -f tflite-android.Dockerfile

inside the Powershell window that has been opened.

Hit Enter and Docker starts downloading TensorFlow latest code, Android SDK and NDK:

After the end of the above procedure copy, paste and run command:

docker run -it -v pwd:/host_dir tf-support-builder bash

to start the container. After some seconds you will see:

Download the tensorflow_lite_support folder

Go to official Google’s repository and download the code. We are specifically interested in the tensorflow_lite_support folder so unzip the file that you have downloaded from the repository and copy paste only the tensorflow_lite_support folder inside the folder that you have the 3 important files.

In case you want to do some changes to the script files or the folder files do them before the below procedure.

Import the script files and tensorflow_lite_support folder at specific locations inside Docker container

To do this procedure we have to open a second Power shell window at the same folder where the scripts are. This is because at the first container we are already in Linux environment inside the container and we cannot move the files from Windows:

We execute command:

docker container ls --all 

to find out the number of containers that exist and their names eg. “d40836790a39”. After that we move the 2 script files and the folder that contains the TensorFlow Lite Support files inside specific locations of the container:
- First execute:

docker cp build_support_aar.sh d40836790a39:/tensorflow_src/tensorflow/lite/tools/

to insert it at “d40836790a39” container and inside “tensorflow_src/tensorflow/lite/tools” folder.

- Second move tensorflow_lite_support folder inside the container by executing:

docker cp tensorflow_lite_support d40836790a39:/tensorflow_src

- Third execute

docker cp build_support_aar_with_docker.sh d40836790a39:/

to move the second script file at the container.

CAUTION: name of the container always change so do not use “d40836790a39" but find out yours and replace at the above 3 commands.

Go back at the first Power Shell window and execute linux commands

First execute:

ls

to view the files.

Second make the script file executable by inserting and running:

chmod +x build_support_aar_with_docker.sh

Third set the location of the Python library by inserting and running:

sudo ln -sf /usr/bin/python3 /usr/bin/python

Finally copy paste:

./build_support_aar_with_docker.sh 

and hit Enter.

Now procedure runs, respond “Yes” to Google’s License agreement and neccessary libraries are downloaded.

After 5–10 minutes, build will be successful.

Move generated folder from container back to Windows folder

At the second Powershell window where we can execute windows commands copy, paste and run

docker cp d40836790a39:tensorflow_src/bazel-bin/tensorflow_lite_support tensorflow-lite-support

With this command (one line command) we move all the generates files from the container back to Windows file system.

CAUTION: name of the container always change so do not use “d40836790a39” but find out yours and replace at the above command.

If we open the tensorflow-lite-support folder and then java folder we can see the generated files. We are specifically interested in tensorflowlite_support.aar file:

Congrats now you have a working .aar file that you can use inside an android studio project!

Use the final .aar file inside Android Studio

Create a libs folder under app’s module folder and insert the .aar file. Out comment the gradle dependency for tensorflow-lite-support and use the code from .aar file the way you see inside the oval red line below:

Colab Notebook

If you use Colab Notebooks you can find the same procedure here. Run the cells one by one and in the last cells select to build Support, Task Vision, Task Text or Task Audio library and C++ targets.

Conclusion

With this procedure you can build a tensorflow_lite_support.aar file from source or you can change the code and create a custom .aar file that will include your custom functions inside. For a demonstration of a custom tensorflow_lite_support.aar file inside an android studio project view a repository here (with_custom_support_library branch). View the builde.gradle and OcrmodelExecutor.kt files to see how to set the .aar file and the usage of the custom library. At this example, conversion to grayscale and getBufferFromOneChannel have been added to the library.

Thanks to Lu Wang and Xunkai Zhang for their help and support. Also special thanks to Thai Nguyen for the original procedure of building TensorFlow Lite with Docker.