Handwritten digits classification from MNIST with TensorFlow on Android; Featuring Tutorial!
Java Python CMake C++
Switch branches/tags
Nothing to show
Clone or download
Latest commit 2fca736 May 13, 2018
Permalink
Failed to load latest commit information.
MnistAndroid Small corrections Apr 21, 2017
images Finished demo Apr 21, 2017
tensorflow_model Improving README and removing MNIST_data Apr 21, 2017
README.md Update README.md May 13, 2018

README.md

MNIST on Android with TensorFlow

Handwritten digits classification from MNIST on Android with TensorFlow.

If you want to make your own version of this app or want to know how to save your model and export it for Android or other devices check the very simple tutorial below.


This work was featured on a video from Siraj Raval!

Check the video demo here.

Image Beautiful art work, right? I know.

How to run this?

Just open this project with Android Studio and is ready to run, this will work with x86 and armeabi-v7a architectures.

How to export my model?

A full example can be seen here

  1. Train your model

  2. Keep an in memory copy of eveything your model learned (like biases and weights) Example: _w = sess.eval(w), where w was learned from training.

  3. Rewrite your model changing the variables for constants with value = in memory copy of learned variables. Example: w_save = tf.constant(_w)

    Also make sure to put names in the input and output of the model, this will be needed for the model later. Example:
    x = tf.placeholder(tf.float32, [None, 1000], name='input')
    y = tf.nn.softmax(tf.matmul(x, w_save) + b_save), name='output')

  4. Export your model with:
    tf.train.write_graph(<graph>, <path for the exported model>, <name of the model>.pb, as_text=False)

How to run my model with Android?

You basically need two things:

  1. The TensorFlow jar
    Move it to the libs folder, right click and add as library.

  2. The TensorFlow so file for the desired architecture:
    x86
    armeabi-v7a

Creat the jniLibs/x86 folder or the jniLibs/armeabi-v7a folder at the main folder.
Move it to app/src/main/jniLibs/x86/libtensorflow_inference.so or app/src/jniLibs/armeabi-v7a/libtensorflow_inference.so

If you want to generate these files yourself, here is a nice tutorial of how to do it.

Interacting with TensorFlow

To interact with TensorFlow you will need an instance of TensorFlowInferenceInterface, you can see more details about it here.

Thank you, have fun!