def run(self):
'''
Build the network and run it.
@rtype: tuple(float32, float32)
@return:
A tuple of: loss, accuracy. Per the test data.
'''
LOG.info("[%s] Doing network run", self._graph.name)
debug = (LOG.getLevelName(LOG.getLogger().level) == 'DEBUG')
with tf.Session(config=tf.ConfigProto(log_device_placement=debug)) as sess:
# Create the network
LOG.info("[%s] Creating the network", self._graph.name)
net_maker = NetMaker(self._graph)
layers = net_maker.make_net()
# Grab the inputs and outputs
in_tensor = layers[ 0]
out_tensor = layers[-1]
# The number of inputs and outputs
num_in = self._x_train.shape[1]
num_out = self._y_train.shape[1]
# Check x_* & y_* and net shapes match
if num_in != in_tensor.shape[1]:
raise ValueError(
"Number of data inputs, %d, "
"does not match network input count, %d, "
"for graph %s" %
(num_in, in_tensor.shape[1], self._graph)
)
if num_out != out_tensor.shape[1]:
raise ValueError(
"Number of data outputs, %d, "
"does not match network output count, %d, "
"for graph %s" %
(num_out, out_tensor.shape[1], self._graph)
)
# Now make the training equipment
truth = tf.placeholder(tf.float32, shape=[None, num_out])
loss = self._create_loss (truth, out_tensor)
accuracy = self._create_accuracy (truth, out_tensor)
optimizer = self._create_optimizer(loss)
# Initialize all variables
sess.run(tf.global_variables_initializer())
# Train for this epoch
LOG.info("[%s] Training the network", self._graph.name)
for epoch in range(1, self._num_epochs + 1):
self._do_epoch(epoch, sess, in_tensor, truth, optimizer)
# And give back the loss and accuracy
LOG.info("[%s] Evaluating the network", self._graph.name)
result = sess.run((loss, accuracy),
feed_dict={ in_tensor : self._x_test,
truth : self._y_test })
LOG.info("[%s] Result: loss=%0.3f accuracy=%0.2f%%",
self._graph.name,
result[0],
100 * result[1])
return result
