# Why is my 1-hidden layer neural network with ReLUs not getting more that 18% accuracy on the notMNIST dataset?

I'm following Google's deep learning course on Udacity. Assignment 2 is to turn a logistic regression network into a 1-hidden layer neural network with rectified linear units and 1024 hidden nodes using Tensorflow.

To do this I've change this:

logits = tf.matmul(tf_train_dataset, weights) + biases


into this:

logits = tf.matmul(tf.nn.relu(tf.matmul(tf_train_dataset, weights1) +
biases1), weights2) + biases2


Here is the full code:

def accuracy(predictions, labels):
return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))
/ predictions.shape[0])

batch_size = 128

graph = tf.Graph()
with graph.as_default():
# Input data. For the training data, we use a placeholder that will be fed
# at run time with a training minibatch.
tf_train_dataset = tf.placeholder(tf.float32,
shape=(batch_size, image_size * image_size))
tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))
tf_valid_dataset = tf.constant(valid_dataset)
tf_test_dataset = tf.constant(test_dataset)

# Variables.
weights1 = tf.Variable(
tf.truncated_normal([image_size * image_size, 1024]))
biases1 = tf.Variable(tf.zeros([1024]))
weights2 = tf.Variable(
tf.truncated_normal([1024, num_labels]))
biases2 = tf.Variable(tf.zeros([num_labels]))

# Training computation.
logits = tf.matmul(tf.nn.relu(tf.matmul(tf_train_dataset, weights1) + biases1), weights2) + biases2
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=tf_train_labels, logits=logits))

# Optimizer.

# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits)
valid_prediction = tf.nn.softmax(
tf.matmul(
tf.nn.relu(
tf.matmul(tf_valid_dataset, weights1)
+ biases1),
weights2) + biases2)
test_prediction = tf.nn.softmax(
tf.matmul(
tf.nn.relu(
tf.matmul(tf_test_dataset, weights1)
+ biases1),
weights2) + biases2)

num_steps = 3001

with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print("Initialized")
for step in range(num_steps):
# Pick an offset within the training data, which has been randomized.
# Note: we could use better randomization across epochs.
offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
# Generate a minibatch.
batch_data = train_dataset[offset:(offset + batch_size), :]
batch_labels = train_labels[offset:(offset + batch_size), :]
# Prepare a dictionary telling the session where to feed the minibatch.
# The key of the dictionary is the placeholder node of the graph to be fed,
# and the value is the numpy array to feed to it.
feed_dict = {tf_train_dataset : batch_data, tf_train_labels : batch_labels}
_, l, predictions = session.run(
[optimizer, loss, train_prediction], feed_dict=feed_dict)
if (step % 500 == 0):
print("Minibatch loss at step %d: %f" % (step, l))
print("Minibatch accuracy: %.1f%%" % accuracy(predictions, batch_labels))
print("Validation accuracy: %.1f%%" % accuracy(
valid_prediction.eval(), valid_labels))
print("Test accuracy: %.1f%%" % accuracy(test_prediction.eval(), test_labels))


Here is the output I get:

Initialized
Minibatch loss at step 0: 208.975021
Minibatch accuracy: 11.7%
Validation accuracy: 10.0%
Minibatch loss at step 500: 0.000000
Minibatch accuracy: 100.0%
Validation accuracy: 10.2%
Minibatch loss at step 1000: 0.000000
Minibatch accuracy: 100.0%
Validation accuracy: 14.6%
Minibatch loss at step 1500: 0.000000
Minibatch accuracy: 100.0%
Validation accuracy: 10.2%
Minibatch loss at step 2000: 0.000000
Minibatch accuracy: 100.0%
Validation accuracy: 17.7%
Minibatch loss at step 2500: 2.952326
Minibatch accuracy: 93.8%
Validation accuracy: 26.6%
Minibatch loss at step 3000: 0.000000
Minibatch accuracy: 100.0%
Validation accuracy: 17.5%
Test accuracy: 18.1%
`

It looks like it's overfitting. It gets close to 100% accuracy on the training data, but only gets around 20% accuracy on the validation and testing data.

I'm confused because I've seen others on Udacity's forum say that they're getting over 90% accuracy using this method. So I think I might be implementing the network incorrectly.

Is this the proper way to implement a 1-hidden layer neural network with rectified linear units? If so, what do you think I'm doing wrong here?

Thank You

• The method by itself doesn't guarantee any particular accuracy level. It depends on your data and how well separated the classes are. – Michael R. Chernick May 2 '17 at 2:27
• @MichaelChernick I'm using the same dataset as the Udacity lesson, the notMNIST dataset. Is there a way that I can check how well separated the classes are without sifting through tens of thousands of images? – Matt D May 2 '17 at 2:30
• What accuracy did they get that was better than yours? – Michael R. Chernick May 2 '17 at 2:31
• I read comments of people getting close to 90% accuracy. Also, what I'm getting now is the same accuracy that I got with a simple linear regression network, which leads me to believe I'm doing something wrong. And the validation accuracy is not increasing with training. – Matt D May 2 '17 at 2:32
• I am not expert enough about your code to tell if you are doing something wrong. But I don't think you know whether the others are doing it right either. – Michael R. Chernick May 2 '17 at 2:35