def model(x, is_train, reuse):
with tf.variable_scope("STN", reuse=reuse):
tl.layers.set_name_reuse(reuse)
nin = InputLayer(x, name='in')
#Localization Network
ln1 = Conv2d(nin, 16, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='tc01')
ln1 = Conv2d(ln1, 32, (3, 3), (2, 2), act=tf.nn.relu, padding='SAME', name='tc02')
ln1 = Conv2d(ln1, 64, (3, 3), (2, 2), act=tf.nn.relu, padding='SAME', name='tc03')
ln1 = BatchNormLayer(ln1, act=tf.nn.relu, is_train=is_train, gamma_init=gamma_init, name='batch_norm1')
ln1 = Conv2d(ln1, 128, (3, 3), (2, 2), act=tf.nn.relu, padding='SAME', name='tc04')
ln1 = BatchNormLayer(ln1, act=tf.nn.relu, is_train=is_train, gamma_init=gamma_init, name='batch_norm2')
ln1 = FlattenLayer(ln1, name='f01')
ln1 = DenseLayer(ln1, n_units=64, act=tf.identity, name='d01')
ln1 = DenseLayer(ln1, n_units=3, act=tf.identity, name='d02')
stn = stn_affine(nin, ln1, out_size=[dim1, dim2], name='ST1')
ln2 = Conv2d(stn, 16, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='tc05')
ln2 = Conv2d(ln2, 32, (3, 3), (2, 2), act=tf.nn.relu, padding='SAME', name='tc06')
ln2 = BatchNormLayer(ln2, act=tf.nn.relu, is_train=is_train, gamma_init=gamma_init, name='batch_norm3')
ln2 = Conv2d(ln2, 64, (3, 3), (2, 2), act=tf.nn.relu, padding='SAME', name='tc07')
ln2 = BatchNormLayer(ln2, act=tf.nn.relu, is_train=is_train, gamma_init=gamma_init, name='batch_norm4')
ln2 = Conv2d(ln2, 128, (2, 2), (2, 2), act=tf.nn.relu, padding='SAME', name='tc08')
ln2 = FlattenLayer(ln2, name='f02')
ln2 = DenseLayer(ln2, n_units=64, act=tf.identity, name='d03')
ln2 = DenseLayer(ln2, n_units=3, act=tf.identity, name='d04')
final_out = tf.add(ln1.outputs,ln2.outputs )
#overall mean error of all the three parameters
ce = tl.cost.mean_squared_error(final_out, y_, 'cost')
L2 = 0
for p in tl.layers.get_variables_with_name('relu/W', True, True):
L2 += tf.contrib.layers.l2_regularizer(0.004)(p)
cost = ce + L2
tf.summary.scalar('cost',cost)
#Individual errors for rotation, x translation and y translation
loss1 = tf.losses.mean_squared_error(final_out[:,0], y_[:,0])
tf.summary.scalar('loss1', loss1)
loss2 = tf.losses.mean_squared_error(final_out[:,1], y_[:,1])
tf.summary.scalar('loss2', loss2)
loss3 = tf.losses.mean_squared_error(final_out[:,2], y_[:,2])
tf.summary.scalar('loss3', loss3)
return ln1,ln2, cost, final_out, loss1, loss2 , loss3, ln1.outputs
def model(x, is_train, reuse):
with tf.variable_scope("STN", reuse=reuse):
tl.layers.set_name_reuse(reuse)
nin = InputLayer(x, name='in')
#Localization Network
ln = Conv2d(nin,
16, (3, 3), (1, 1),
act=tf.nn.relu,
padding='SAME',
name='tc01')
ln = Conv2d(nin,
64, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc02')
ln = Conv2d(ln,
128, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc03')
ln = FlattenLayer(ln, name='f01')
ln = DenseLayer(ln, n_units=64, act=tf.identity, name='d01')
ln = DenseLayer(ln, n_units=3, act=tf.identity, name='d02')
#STN
stn = stn_affine(nin, ln, out_size=[dim1, dim2], name='ST')
nt = Conv2d(stn,
16, (3, 3), (1, 1),
act=tf.nn.relu,
padding='SAME',
name='tc1')
nt = Conv2d(nt,
32, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc2')
# 32x32x16
nt = Conv2d(nt,
64, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc3')
nt = Conv2d(nt,
128, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc4')
nt = FlattenLayer(nt, name='f')
nt = DenseLayer(nt, n_units=64, act=tf.identity, name='d1')
nt = DenseLayer(nt, n_units=3, act=tf.identity, name='do')
y = nt.outputs
localization_out = ln.outputs
final_out = y #tf.add(localization_out, y)
#overall mean error of all the three parameters
cost = tl.cost.mean_squared_error(final_out, y_, 'cost')
tf.summary.scalar('cost', cost)
#Individual errors for rotation, x translation and y translation
loss1 = tf.losses.mean_squared_error(final_out[:, 0], y_[:, 0])
tf.summary.scalar('loss1', loss1)
loss2 = tf.losses.mean_squared_error(final_out[:, 1], y_[:, 1])
tf.summary.scalar('loss2', loss2)
loss3 = tf.losses.mean_squared_error(final_out[:, 2], y_[:, 2])
tf.summary.scalar('loss3', loss3)
return nt, cost, final_out, loss1, loss2, loss3, localization_out
def model(x, is_train, reuse):
with tf.variable_scope("STN", reuse=reuse):
tl.layers.set_name_reuse(reuse)
nin = InputLayer(x, name='in')
#Localization Network
ln1 = Conv2d(nin,
16, (3, 3), (1, 1),
act=tf.nn.relu,
padding='SAME',
name='tc01')
ln1 = Conv2d(ln1,
64, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc02')
ln1 = Conv2d(ln1,
128, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc03')
ln1 = FlattenLayer(ln1, name='f01')
ln1 = DropoutLayer(ln1, 0.5, True, is_train, name='drop1')
ln1 = DenseLayer(ln1, n_units=64, act=tf.identity, name='d01')
ln1 = DropoutLayer(ln1, 0.5, True, is_train, name='drop2')
ln1 = DenseLayer(ln1, n_units=3, act=tf.identity, name='d02')
stn = stn_affine(nin, ln1, out_size=[dim1, dim2], name='ST1')
ln2 = Conv2d(stn,
16, (3, 3), (1, 1),
act=tf.nn.relu,
padding='SAME',
name='tc04')
ln2 = Conv2d(ln2,
64, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc05')
ln2 = Conv2d(ln2,
128, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc06')
ln2 = FlattenLayer(ln2, name='f02')
ln2 = DropoutLayer(ln2, 0.5, True, is_train, name='drop3')
ln2 = DenseLayer(ln2, n_units=64, act=tf.identity, name='d03')
ln2 = DropoutLayer(ln2, 0.5, True, is_train, name='drop4')
ln2 = DenseLayer(ln2, n_units=3, act=tf.identity, name='d04')
stn2 = stn_affine(stn, ln2, out_size=[dim1, dim2], name='ST2')
ln3 = Conv2d(stn2,
16, (3, 3), (1, 1),
act=tf.nn.relu,
padding='SAME',
name='tc07')
ln3 = Conv2d(ln3,
64, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc08')
ln3 = Conv2d(ln3,
128, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc09')
ln3 = FlattenLayer(ln3, name='f03')
ln3 = DropoutLayer(ln3, 0.5, True, is_train, name='drop5')
ln3 = DenseLayer(ln3, n_units=64, act=tf.identity, name='d05')
ln3 = DropoutLayer(ln3, 0.5, True, is_train, name='drop6')
ln3 = DenseLayer(ln3, n_units=3, act=tf.identity, name='d06')
stn3 = stn_affine(stn2, ln3, out_size=[dim1, dim2], name='ST3')
ln4 = Conv2d(stn3,
16, (3, 3), (1, 1),
act=tf.nn.relu,
padding='SAME',
name='tc10')
ln4 = Conv2d(ln4,
64, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc11')
ln4 = Conv2d(ln4,
128, (3, 3), (2, 2),
act=tf.nn.relu,
padding='SAME',
name='tc12')
ln4 = FlattenLayer(ln4, name='f04')
ln4 = DropoutLayer(ln4, 0.5, True, is_train, name='drop7')
ln4 = DenseLayer(ln4, n_units=64, act=tf.identity, name='d07')
ln4 = DropoutLayer(ln4, 0.5, True, is_train, name='drop8')
ln4 = DenseLayer(ln4, n_units=3, act=tf.identity, name='d08')
final_out = tf.add(ln1.outputs, ln2.outputs)
final_out = tf.add(final_out, ln3.outputs)
final_out = tf.add(final_out, ln4.outputs)
#overall mean error of all the three parameters
cost = tl.cost.mean_squared_error(final_out, y_, 'cost')
tf.summary.scalar('cost', cost)
#Individual errors for rotation, x translation and y translation
loss1 = tf.losses.mean_squared_error(final_out[:, 0], y_[:, 0])
tf.summary.scalar('loss1', loss1)
loss2 = tf.losses.mean_squared_error(final_out[:, 1], y_[:, 1])
tf.summary.scalar('loss2', loss2)
loss3 = tf.losses.mean_squared_error(final_out[:, 2], y_[:, 2])
tf.summary.scalar('loss3', loss3)
return ln1, ln2, ln3, ln4, cost, final_out, loss1, loss2, loss3, ln1.outputs