I am following the third Jupyter notebook on Tensorflow examples.
Running problem 4, I tried to implement a function which builds automatically a number of hidden layers, without manually coding the configuration of each layer.
However, the model runs providing very low accuracy (10%) so I thought that maybe such function could not be compatible with the graph builder of Tensorflow.
My code is the following:
def hlayers(n_layers, n_nodes, i_size, a, r=0, keep_p=1):
for i in range(n_layers):
if i > 0:
i_size = n_nodes
w = tf.Variable(tf.truncated_normal([i_size, n_nodes]), name=f'W{i}')
b = tf.Variable(tf.zeros([n_nodes]), name=f'b{i}')
pa = tf.nn.relu(tf.add(tf.matmul(a, w), b))
a = tf.nn.dropout(pa, keep_prob=keep_p, name=f'a{i}')
r += tf.nn.l2_loss(w, name=f'r{i}')
return a, r
batch_size = 128
num_nodes = 1024
beta = 0.01
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),
name='Dataset')
tf_train_labels = tf.placeholder(
tf.float32,
shape=(batch_size, num_labels),
name='Labels')
tf_valid_dataset = tf.constant(valid_dataset)
tf_test_dataset = tf.constant(test_dataset)
keep_p = tf.placeholder(tf.float32, name='KeepProb')
# Hidden layers.
a, r = hlayers(
n_layers=3,
n_nodes=num_nodes,
i_size=image_size * image_size,
a=tf_train_dataset,
keep_p=keep_p)
# Output layer.
wo = tf.Variable(tf.truncated_normal([num_nodes, num_labels]), name='Wo')
bo = tf.Variable(tf.zeros([num_labels]), name='bo')
logits = tf.add(tf.matmul(a, wo), bo, name='Logits')
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(
labels=tf_train_labels, logits=logits))
# Regularizer.
regularizers = tf.add(r, tf.nn.l2_loss(wo))
loss = tf.reduce_mean(loss + beta * regularizers, name='Loss')
# Optimizer.
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits)
a, _ = hlayers(
n_layers=3,
n_nodes=num_nodes,
i_size=image_size * image_size,
a=tf_valid_dataset)
valid_prediction = tf.nn.softmax(tf.add(tf.matmul(a, wo), bo))
a, _ = hlayers(
n_layers=3,
n_nodes=num_nodes,
i_size=image_size * image_size,
a=tf_test_dataset)
test_prediction = tf.nn.softmax(tf.add(tf.matmul(a, wo), bo))
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,
keep_p : 0.5}
_, 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))
