Creating Custom Model For Android Using TensorFlow

As I had promised in my previous article on building TensorFlow for Android that I will be writing an article on How to train custom model for Android using TensorFlow. So, I have written this article. Still more to come.

If you have not checked my article on building TensorFlow for Android, check here.

In this article, we will train a model to recognize the handwritten digits. Here, we will use the famous MNIST Image Dataset which like the Hello World in Machine Learning for simplicity.

Credit: The classifier example has been taken from Google TensorFlow example. The custom drawing view used in this project is taken from here.

Here is the complete sample project.

The complete code to train and save the model which will directly run on Android is below. Read it carefully.

from __future__ import print_function
import shutil
import os.path
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

EXPORT_DIR = './model'

if os.path.exists(EXPORT_DIR):
    shutil.rmtree(EXPORT_DIR)

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

# Parameters
learning_rate = 0.001
training_iters = 200000
batch_size = 128
display_step = 10

# Network Parameters
n_input = 784  # MNIST data input (img shape: 28*28)
n_classes = 10  # MNIST total classes (0-9 digits)
dropout = 0.75  # Dropout, probability to keep units

# tf Graph input
x = tf.placeholder(tf.float32, [None, n_input])
y = tf.placeholder(tf.float32, [None, n_classes])
keep_prob = tf.placeholder(tf.float32)  # dropout (keep probability)


# Create some wrappers for simplicity
def conv2d(x, W, b, strides=1):
    # Conv2D wrapper, with bias and relu activation
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)


def maxpool2d(x, k=2):
    # MaxPool2D wrapper
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
                          padding='SAME')


# Create Model
def conv_net(x, weights, biases, dropout):
    # Reshape input picture
    x = tf.reshape(x, shape=[-1, 28, 28, 1])

    # Convolution Layer
    conv1 = conv2d(x, weights['wc1'], biases['bc1'])
    # Max Pooling (down-sampling)
    conv1 = maxpool2d(conv1, k=2)

    # Convolution Layer
    conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
    # Max Pooling (down-sampling)
    conv2 = maxpool2d(conv2, k=2)

    # Fully connected layer
    # Reshape conv2 output to fit fully connected layer input
    fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])
    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
    fc1 = tf.nn.relu(fc1)
    # Apply Dropout
    fc1 = tf.nn.dropout(fc1, dropout)

    # Output, class prediction
    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])
    return out


# Store layers weight & bias
weights = {
    # 5x5 conv, 1 input, 32 outputs
    'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),
    # 5x5 conv, 32 inputs, 64 outputs
    'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64])),
    # fully connected, 7*7*64 inputs, 1024 outputs
    'wd1': tf.Variable(tf.random_normal([7 * 7 * 64, 1024])),
    # 1024 inputs, 10 outputs (class prediction)
    'out': tf.Variable(tf.random_normal([1024, n_classes]))
}

biases = {
    'bc1': tf.Variable(tf.random_normal([32])),
    'bc2': tf.Variable(tf.random_normal([64])),
    'bd1': tf.Variable(tf.random_normal([1024])),
    'out': tf.Variable(tf.random_normal([n_classes]))
}

# Construct model
pred = conv_net(x, weights, biases, keep_prob)

# Define loss and optimizer
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

# Evaluate model
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

# Initializing the variables
init = tf.initialize_all_variables()

# Launch the graph
with tf.Session() as sess:
    sess.run(init)
    step = 1
    # Keep training until reach max iterations
    while step * batch_size < training_iters:
        batch_x, batch_y = mnist.train.next_batch(batch_size)
        # Run optimization op (backprop)
        sess.run(optimizer, feed_dict={x: batch_x, y: batch_y,
                                       keep_prob: dropout})
        if step % display_step == 0:
            # Calculate batch loss and accuracy
            loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x,
                                                              y: batch_y,
                                                              keep_prob: 1.})
            print("Iter " + str(step * batch_size) + ", Minibatch Loss= " + \
                  "{:.6f}".format(loss) + ", Training Accuracy= " + \
                  "{:.5f}".format(acc))
        step += 1
    print("Optimization Finished!")

    # Calculate accuracy for 256 mnist test images
    print("Testing Accuracy:", \
          sess.run(accuracy, feed_dict={x: mnist.test.images[:256],
                                        y: mnist.test.labels[:256],
                                        keep_prob: 1.}))
    WC1 = weights['wc1'].eval(sess)
    BC1 = biases['bc1'].eval(sess)
    WC2 = weights['wc2'].eval(sess)
    BC2 = biases['bc2'].eval(sess)
    WD1 = weights['wd1'].eval(sess)
    BD1 = biases['bd1'].eval(sess)
    W_OUT = weights['out'].eval(sess)
    B_OUT = biases['out'].eval(sess)

# Create new graph for exporting
g = tf.Graph()
with g.as_default():
    x_2 = tf.placeholder("float", shape=[None, 784], name="input")

    WC1 = tf.constant(WC1, name="WC1")
    BC1 = tf.constant(BC1, name="BC1")
    x_image = tf.reshape(x_2, [-1, 28, 28, 1])
    CONV1 = conv2d(x_image, WC1, BC1)
    MAXPOOL1 = maxpool2d(CONV1, k=2)

    WC2 = tf.constant(WC2, name="WC2")
    BC2 = tf.constant(BC2, name="BC2")
    CONV2 = conv2d(MAXPOOL1, WC2, BC2)
    MAXPOOL2 = maxpool2d(CONV2, k=2)

    WD1 = tf.constant(WD1, name="WD1")
    BD1 = tf.constant(BD1, name="BD1")

    FC1 = tf.reshape(MAXPOOL2, [-1, WD1.get_shape().as_list()[0]])
    FC1 = tf.add(tf.matmul(FC1, WD1), BD1)
    FC1 = tf.nn.relu(FC1)

    W_OUT = tf.constant(W_OUT, name="W_OUT")
    B_OUT = tf.constant(B_OUT, name="B_OUT")

    # skipped dropout for exported graph as there 
    # is no need for already calculated weights

    OUTPUT = tf.nn.softmax(tf.matmul(FC1, W_OUT) + B_OUT, name="output")

    sess = tf.Session()
    init = tf.initialize_all_variables()
    sess.run(init)

    graph_def = g.as_graph_def()
    tf.train.write_graph(graph_def, EXPORT_DIR, 'mnist_model_graph.pb', as_text=False)

    # Test trained model
    y_train = tf.placeholder("float", [None, 10])
    correct_prediction = tf.equal(tf.argmax(OUTPUT, 1), tf.argmax(y_train, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

    print("check accuracy %g" % accuracy.eval(
            {x_2: mnist.test.images, y_train: mnist.test.labels}, sess))

Explanation of code:

  • First, it downloads the dataset from the MNIST.
  • Then it starts training the model.
  • After training, it checks accuracy.
  • Then, it saves the model to the given path and checks the accuracy again.

Check the complete python code here.

It is very easy to understand it by reading the python code.

When we run this code, we get the required trained model that will be used in the Android application.

If you are getting any problem in building the project, connect with me, I will be happy to help.

Happy Coding :)

Learn Machine Learning with Tensorflow from here.

Update : Check Android TensorFlow Lite Machine Learning Example

Also, Let’s become friends on Twitter, Linkedin, Github, and Facebook.