Neural Network with Eager API

Build a 2-hidden layers fully connected neural network (a.k.a multilayer perceptron) with TensorFlow's Eager API.

This example is using some of TensorFlow higher-level wrappers (tf.estimators, tf.layers, tf.metrics, ...), you can check 'neural_network_raw' example for a raw, and more detailed TensorFlow implementation.

• Author: Aymeric Damien
• Project: https://github.com/aymericdamien/TensorFlow-Examples/

Neural Network Overview

MNIST Dataset Overview

This example is using MNIST handwritten digits. The dataset contains 60,000 examples for training and 10,000 examples for testing. The digits have been size-normalized and centered in a fixed-size image (28x28 pixels) with values from 0 to 1. For simplicity, each image has been flattened and converted to a 1-D numpy array of 784 features (28*28).

More info: http://yann.lecun.com/exdb/mnist/

from __future__ import print_function

import tensorflow as tf
import tensorflow.contrib.eager as tfe

# Set Eager API
tfe.enable_eager_execution()

# Import MNIST data
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=False)

Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz

# Parameters
learning_rate = 0.001
num_steps = 1000
batch_size = 128
display_step = 100

# Network Parameters
n_hidden_1 = 256 # 1st layer number of neurons
n_hidden_2 = 256 # 2nd layer number of neurons
num_input = 784 # MNIST data input (img shape: 28*28)
num_classes = 10 # MNIST total classes (0-9 digits)

# Using TF Dataset to split data into batches
dataset = tf.data.Dataset.from_tensor_slices(
    (mnist.train.images, mnist.train.labels)).batch(batch_size)
dataset_iter = tfe.Iterator(dataset)

# Define the neural network. To use eager API and tf.layers API together,
# we must instantiate a tfe.Network class as follow:
class NeuralNet(tfe.Network):
    def __init__(self):
        # Define each layer
        super(NeuralNet, self).__init__()
        # Hidden fully connected layer with 256 neurons
        self.layer1 = self.track_layer(
            tf.layers.Dense(n_hidden_1, activation=tf.nn.relu))
        # Hidden fully connected layer with 256 neurons
        self.layer2 = self.track_layer(
            tf.layers.Dense(n_hidden_2, activation=tf.nn.relu))
        # Output fully connected layer with a neuron for each class
        self.out_layer = self.track_layer(tf.layers.Dense(num_classes))

    def call(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        return self.out_layer(x)


neural_net = NeuralNet()

# Cross-Entropy loss function
def loss_fn(inference_fn, inputs, labels):
    # Using sparse_softmax cross entropy
    return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
        logits=inference_fn(inputs), labels=labels))


# Calculate accuracy
def accuracy_fn(inference_fn, inputs, labels):
    prediction = tf.nn.softmax(inference_fn(inputs))
    correct_pred = tf.equal(tf.argmax(prediction, 1), labels)
    return tf.reduce_mean(tf.cast(correct_pred, tf.float32))


# SGD Optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)

# Compute gradients
grad = tfe.implicit_gradients(loss_fn)

# Training
average_loss = 0.
average_acc = 0.
for step in range(num_steps):

    # Iterate through the dataset
    try:
        d = dataset_iter.next()
    except StopIteration:
        # Refill queue
        dataset_iter = tfe.Iterator(dataset)
        d = dataset_iter.next()

    # Images
    x_batch = d[0]
    # Labels
    y_batch = tf.cast(d[1], dtype=tf.int64)

    # Compute the batch loss
    batch_loss = loss_fn(neural_net, x_batch, y_batch)
    average_loss += batch_loss
    # Compute the batch accuracy
    batch_accuracy = accuracy_fn(neural_net, x_batch, y_batch)
    average_acc += batch_accuracy

    if step == 0:
        # Display the initial cost, before optimizing
        print("Initial loss= {:.9f}".format(average_loss))

    # Update the variables following gradients info
    optimizer.apply_gradients(grad(neural_net, x_batch, y_batch))

    # Display info
    if (step + 1) % display_step == 0 or step == 0:
        if step > 0:
            average_loss /= display_step
            average_acc /= display_step
        print("Step:", '%04d' % (step + 1), " loss=",
              "{:.9f}".format(average_loss), " accuracy=",
              "{:.4f}".format(average_acc))
        average_loss = 0.
        average_acc = 0.

Initial loss= 2.340397596
Step: 0001  loss= 2.340397596  accuracy= 0.0703
Step: 0100  loss= 0.586046159  accuracy= 0.8305
Step: 0200  loss= 0.253318846  accuracy= 0.9282
Step: 0300  loss= 0.214748293  accuracy= 0.9377
Step: 0400  loss= 0.180644721  accuracy= 0.9466
Step: 0500  loss= 0.137285724  accuracy= 0.9591
Step: 0600  loss= 0.119845696  accuracy= 0.9636
Step: 0700  loss= 0.113618039  accuracy= 0.9665
Step: 0800  loss= 0.109642141  accuracy= 0.9676
Step: 0900  loss= 0.085067607  accuracy= 0.9746
Step: 1000  loss= 0.079819344  accuracy= 0.9754

# Evaluate model on the test image set
testX = mnist.test.images
testY = mnist.test.labels

test_acc = accuracy_fn(neural_net, testX, testY)
print("Testset Accuracy: {:.4f}".format(test_acc))

Testset Accuracy: 0.9719