This is the code for"A Guide to Running Tensorflow Models on Android" By SIraj Raval on Youtube

Jupyter Notebook 0.69% Python 1.42% Java 97.89%

a_guide_to_running_tensorflow_models_on_android's Introduction

A_Guide_to_Running_Tensorflow_Models_on_Android

This is the code for"A Guide to Running Tensorflow Models on Android" By SIraj Raval on Youtube

Overview

This is the code for this video on Youtube by Siraj Raval.

Handwritten digits classification from MNIST on Android with TensorFlow.

If you want to make your own version of this app or want to knowhow to save your model and export it for Android or other devices check the very simple tutorial below. The UI and expert-graph.pb model were taken from: https://github.com/miyosuda/TensorFlowAndroidMNIST, so thank you miyousuda.

Dependencies

All included

Usage

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

Train your model
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.
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')
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 need tensorflow.aar, which can be downloaded from the nightly build artifact of TensorFlow CI, here we use the #124 build.

Interacting with TensorFlow

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

Credits

Credits for this code go to mari-linhares. I've merely created a wrapper to get people started.

a_guide_to_running_tensorflow_models_on_android's People

Contributors

Stargazers

Watchers

Forkers

zahorecztibor jdc08161063 dconstan stevenlol nieshaoshuai androiddream karanchaparwal rcpbayindir grantad aiwebbot huleg neuralnetworkingtechnologies mkodekar antsmithdev saurabhsingh13no mano42023 huabawa sagarnildass annie1ustc guruprasaad123 ebrgh12 alinasser96 ishandutta2007 akssieg piotrbazan minosai jjjjohnson tracyghz bigfoxses shahariarrabby mohitkhatwani csearman lorepieri8 shubhampachori12110095 aravindsasidharan neo4reo rohit-gupta-web3 inzkhan anmolduainter woshiwanglitao maxpastor rajap144 hehesky ynb97 tzebin omaraflak pandamic belal-bh circlez3791117 poentino87 zizo216 rohitag13 shaitender-intg ankitdamani abinfinity juanlp m3bin hanfeijp naidubharadwaj9 jags-aaron juriigen abhinavraj23 septiandrd habibzadeh bambanghariantosianturi meto16 stella-gao vinayprabhu19 jiaoma swap1505 pedrohfp gaoleia pramodbharti humorbei rakeshjha4385 andresmaca huibaobao kanavanand ghayasahmed786 harshitsoni97 abhishek-07 bantamjoe amanksingh007 anhvaut dzaiats sidcgithub devangpaliwal wyjulia eong2012 unclesugar sagarsishir51 faisalomar rh01 gscse givaldofrancisco mrhubgit stephanzmk wasim6691 pipesanta marykravets

a_guide_to_running_tensorflow_models_on_android's Issues

pb file protection for deployment

Hello, this is not an issue well, it may be for anyone publishing an app with a .pb file in it and wanting to prevent unauthorized usage of such pb file.

=> Could you please help us figuring out how to decrypt on runtime & successfully load our pb file on Android?

The idea would be to include the encrypted pb in the apk and store the key as safe as possible (+ preventing the app from running on rooted / debuggin devices, etc)

Please help us if you have an idea on how to achieve this; we found libraries to encrypt / decrypt text files in Android, but we are missing the decryption into pb / loading the decrypted model step....

Load test data from file

Hi,
Is it possible to generate an alternative app activity to load the test data from a file(csv/image) and test the model's performance?
I am working on an inertial sensor(acc+gyro) based app for activity classification would like to have some suggestion on an offline testing from a loaded test data file.

Many thanks.

import dependency

from tensorflow.python.framework.graph_util import convert_variables_to_constants

Convert variables to constants and save

output_graph_def = convert_variables_to_constants(sess, sess.graph_def, output_node_names=
['output/outputs'])
    with tf.gfile.FastGFile('model/CTNModel.pb', mode='wb') as f:
        f.write(output_graph_def.SerializeToString())

Binary Output

export_model(tf.train.Saver(), model, ["conv2d_1_input"], "dense_2/Softmax")

In the above line of code, what would be the parameters for Binary output?
I have used only ONE convolution layer.

I just want to make everything clear before training.
Thanking you in anticipation.

llsourcell / a_guide_to_running_tensorflow_models_on_android Goto Github PK