Validation and Test with TensorFlow

Question

I have created a one hidden layer neural network with a pyramidal architecture using TensorFlow. Here is the code:

num_classes = 10
image_size = 28

#Read the data
train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels = OpenDataSets("...")
#Create and convert what is needed.
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)

#Then I create the NN.
Wh = tf.Variable(tf.truncated_normal([image_size * image_size, image_size * image_size / 2]))
bh = tf.Variable(tf.truncated_normal([image_size * image_size / 2]))
hidden = tf.nn.relu(tf.matmul(tf_train_dataset, Wh) + bh)

Wout = tf.Variable(tf.truncated_normal([image_size * image_size / 2, num_labels]))
bout = tf.Variable(tf.truncated_normal([num_labels]))
logits = tf.nn.relu(tf.matmul(hidden, Wout) + bout)

loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
train_prediction = tf.nn.softmax(logits)

And now I train my NN:

with tf.Session(graph=graph) as session:
    tf.initialize_all_variables().run()
    for step in range(1000):
        offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
        batch_data = train_dataset[offset:(offset + batch_size), :]
        batch_labels = train_labels[offset:(offset + batch_size), :]
        feed_dict = {tf_train_dataset : batch_data, tf_train_labels : batch_labels}
        _, l, predictions = session.run([optimizer, loss, train_prediction], feed_dict=feed_dict)

Now I would like to validate and test my NN after training. But I don't know how to create the new feed_dict and use session.run in order to validate/test.

Thanks for your help!

Answer 1

You will first have to create appropriate validation/test tensor functions. For one-layer MPL, it involves nested multiply with weights and addition of biases (and also Relu's since you have them in your original model). Define these right below your train predictions

valid_prediction = tf.nn.softmax(
                      tf.nn.relu(tf.matmul(
                         tf.nn.relu(tf.matmul(tf_valid_dataset, Wh) + bh)), Wout) + bout)))
test_prediction = tf.nn.softmax(
                      tf.nn.relu(tf.matmul(
                         tf.nn.relu(tf.matmul(tf_test_dataset, Wh) + bh)), Wout) + bout)))

These expressions are in fact quite identical to logit variable defined in your code, only using tf_valid_dataset and tf_test_dataset respectively. You can create intermediate variables to simplify them.

You will then have to create some validation/test function to test accuracy. Simplest would be to test most likely predicted class (Misclassification error roughly). Define this outside your graph/session.

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

After that, you can simply pass this accuracy function inside same session/feed_dict to compute validation/test score.

print 'Validation accuracy: %.1f%%' % accuracy(valid_prediction.eval(), valid_labels)
print 'Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels)