def withGrowth():
global train_x, train_y, max_iterations, batch_size
with tf.Session() as session:
net = Net(session, INPUT_DIM, OUTPUT_DIM)
net = net.add_hidden_layer(session,
1,
bactivate=True,
non_liniarity=tf.nn.sigmoid)
net = net.add_hidden_layer(session,
1,
bactivate=True,
non_liniarity=tf.nn.sigmoid)
last_loss = 1000000000.0
loss_counts = 0
for i in range(max_iterations - no_back_iterations):
train_x, train_y = shuffle(train_x, train_y)
loss = net.train(train_x, train_y, batch_size=batch_size)
print(i, loss)
if loss > last_loss:
if loss_counts > 9:
print "adding new nodes"
back_loss = net.get_reconstruction_error_per_hidden_layer(
train_x, train_y)
print "Back loss %s" % (back_loss, )
layer_with_greatest_back_loss = back_loss.index(
max(back_loss))
net = net.add_node_to_hidden_layer(
session, layer_with_greatest_back_loss)
print net.hidden_nodes
last_loss = 10000000000.0
loss_counts = 0
else:
last_loss = loss
loss_counts += 1
else:
last_loss = loss
net.use_bactivate = False
for j in range(no_back_iterations):
i = j + max_iterations - no_back_iterations
train_x, train_y = shuffle(train_x, train_y)
loss = net.train(train_x, train_y, batch_size=batch_size)
print(i, loss)
print "final loss %s, %s" % (i, loss)
print "nodes %s" % (net.hidden_nodes, )
def withGrowth():
global train_x, train_y, max_iterations, batch_size
with tf.Session() as session:
net = Net(session, INPUT_DIM, OUTPUT_DIM)
net = net.add_hidden_layer(session, 1, bactivate=True, non_liniarity=tf.nn.sigmoid)
net = net.add_hidden_layer(session, 1, bactivate=True, non_liniarity=tf.nn.sigmoid)
last_loss = 1000000000.0
loss_counts = 0
for i in range(max_iterations - no_back_iterations):
train_x, train_y = shuffle(train_x, train_y)
loss = net.train(train_x, train_y, batch_size=batch_size)
print(i, loss)
if loss > last_loss:
if loss_counts > 9:
print "adding new nodes"
back_loss = net.get_reconstruction_error_per_hidden_layer(train_x, train_y)
print "Back loss %s" % (back_loss,)
layer_with_greatest_back_loss = back_loss.index(max(back_loss))
net = net.add_node_to_hidden_layer(session, layer_with_greatest_back_loss)
print net.hidden_nodes
last_loss = 10000000000.0
loss_counts = 0
else:
last_loss = loss
loss_counts += 1
else:
last_loss = loss
net.use_bactivate = False
for j in range(no_back_iterations):
i = j + max_iterations - no_back_iterations
train_x, train_y = shuffle(train_x, train_y)
loss = net.train(train_x, train_y, batch_size=batch_size)
print(i, loss)
print "final loss %s, %s" % (i, loss)
print "nodes %s" % (net.hidden_nodes, )
def noGrowth(layers, bactivate=False):
global train_x, train_y, max_iterations, batch_size
with tf.Session() as session:
net = Net(session, INPUT_DIM, OUTPUT_DIM)
for l in layers:
net = net.add_hidden_layer(session, l, bactivate=bactivate)
for i in range(max_iterations):
train_x, train_y = shuffle(train_x, train_y)
loss = net.train(train_x, train_y, batch_size=batch_size)
print(i, loss)
print "final loss %s, %s" % (i, loss)
return loss
def noGrowth(layers, bactivate=False):
global train_x, train_y, max_iterations, batch_size
with tf.Session() as session:
net = Net(session, INPUT_DIM, OUTPUT_DIM)
for l in layers:
net = net.add_hidden_layer(session, l, bactivate=bactivate)
for i in range(max_iterations):
train_x, train_y = shuffle(train_x, train_y)
loss = net.train(train_x, train_y, batch_size=batch_size)
print(i, loss)
print "final loss %s, %s" % (i, loss)
return loss