def kpblock6dn_elu(x3D, batch_size, k, name="kpblock6dn"):
num = 4 * k
shape = [[6, num // 2], [num // 2, num // 4], [num // 4, num // 8],
[1, num // 8], [num // 8, 3], [1, 3]]
param = []
init = tf.glorot_normal_initializer()
reg = tf.contrib.layers.l2_regularizer(1e-4)
for i in range(len(shape)):
param.append(
tf.get_variable(shape=shape[i],
regularizer=reg,
name=name + 'param_%d' % i))
_, knn_index = group.knn(x3D, k)
x3Dkcnnglobal = tf.gather_nd(x3D, knn_index, name=name + "_gather")
x3Dkcnnlocal = x3Dkcnnglobal - tf.reshape(x3D, [batch_size, -1, 1, 3])
x3Dkcnn = tf.concat([x3Dkcnnglobal, x3Dkcnnlocal], 3)
x3Dkcnn = tf.reshape(x3Dkcnn, [-1, 6])
x3Dkcnn = tf.matmul(x3Dkcnn, param[0])
x3Dkcnn = tf.nn.elu(x3Dkcnn)
x3Dkcnn = tf.matmul(x3Dkcnn, param[1])
x3Dkcnn = tf.reshape(x3Dkcnn, [-1, k, num // 4])
x3Dkcnn = tf.reduce_max(x3Dkcnn, 1)
x3Dkcnn = tf.matmul(x3Dkcnn, param[2]) + param[3]
x3Dkcnn = tf.nn.elu(x3Dkcnn)
x3Dkcnn = tf.matmul(x3Dkcnn, param[4]) + param[5]
x3Dkcnn = tf.reshape(x3Dkcnn, [batch_size, -1, 3])
x3D += x3Dkcnn
return x3D
def kpsmooth(x3D, batch_size, k, name="kpsmooth"):
_, knn_index = group.knn(x3D, k)
x3Dkcnnglobal = tf.gather_nd(x3D, knn_index, name=name + "_gather")
x3Dkcnnglobal = tf.concat(
[x3Dkcnnglobal, tf.reshape(x3D, [batch_size, -1, 1, 3])], axis=2)
return tf.reduce_mean(x3Dkcnnglobal, axis=2)
def kpblock6ext(x2D, x3D, batch_size, k, name="kpblock6", reuse=False):
x2D = conv_2d(x2D,
16, (3, 3),
strides=1,
activation='relu',
weight_decay=1e-5,
regularizer='L2',
name=name + "00",
reuse=reuse)
x2D = conv_2d(x2D,
8, (3, 3),
strides=1,
activation='relu',
weight_decay=1e-5,
regularizer='L2',
name=name + "01",
reuse=reuse)
x2Da = conv_2d(x2D,
1, (1, 1),
strides=1,
activation='linear',
weight_decay=1e-5,
regularizer='L2',
name=name + "02",
reuse=reuse)
x2Da = tf.nn.max_pool(x2Da, [1, 2, 2, 1], [1, 2, 2, 1], "VALID")
x2Db = tf.nn.max_pool(x2D, [1, 2, 2, 1], [1, 2, 2, 1], "VALID")
x2Db = conv_2d(x2Db,
1, (1, 1),
strides=1,
activation='linear',
weight_decay=1e-5,
regularizer='L2',
name=name + "03",
reuse=reuse)
x2D = tf.nn.relu(x2Da + x2Db)
x2Ddim = int(x2D.shape[1]) * int(x2D.shape[2]) * int(x2D.shape[3])
num = 4 * k
n = [
6 * (num // 2), (num // 2) * (num // 4), (num // 4) * (num // 8),
(num // 8), (num // 8) * 3, 3
]
shape = [[batch_size, 6, num // 2], [batch_size, num // 2, num // 4],
[batch_size, num // 4, num // 8], [batch_size, 1, num // 8],
[batch_size, num // 8, 3], [batch_size, 1, 3]]
param = []
fc_scope = tf.variable_scope(name, reuse=reuse)
for i in range(len(n)):
if 2 * n[i] < x2Ddim:
x1D = tflearn.layers.core.fully_connected(x2D,
2 * n[i],
activation='relu',
weight_decay=1e-4,
regularizer='L2',
reuse=reuse,
scope=fc_scope)
else:
x1D = tflearn.layers.core.fully_connected(x2D,
n[i],
activation='relu',
weight_decay=1e-4,
regularizer='L2',
reuse=reuse,
scope=fc_scope)
x1D = tflearn.layers.core.fully_connected(x1D,
n[i],
activation='linear',
weight_decay=1e-4,
regularizer='L2',
reuse=reuse,
scope=fc_scope)
param.append(tf.reshape(x1D, shape[i]))
_, knn_index = group.knn(x3D, k)
x3Dkcnnglobal = tf.gather_nd(x3D, knn_index, name=name + "_gather")
x3Dkcnnlocal = x3Dkcnnglobal - tf.reshape(x3D, [batch_size, -1, 1, 3])
x3Dkcnn = tf.concat([x3Dkcnnglobal, x3Dkcnnlocal], 3)
x3Dkcnn = tf.reshape(x3Dkcnn, [batch_size, -1, 6])
x3Dkcnn = tf.matmul(x3Dkcnn, param[0])
x3Dkcnn = tf.nn.relu(x3Dkcnn)
x3Dkcnn = tf.matmul(x3Dkcnn, param[1])
x3Dkcnn = tf.reshape(x3Dkcnn, [batch_size, -1, k, num // 4])
x3Dkcnn = tf.reduce_max(x3Dkcnn, 2)
x3Dkcnn = tf.matmul(x3Dkcnn, param[2]) + param[3]
x3Dkcnn = tf.nn.relu(x3Dkcnn)
x3Dkcnn = tf.matmul(x3Dkcnn, param[4]) + param[5]
x3D += x3Dkcnn
return x3D