def reconstructor_tf(x, angles, reuse = True):
with tf.variable_scope("Reconstructor", reuse = reuse):
aco = tf.cos(angles*math.pi)
asi = tf.sin(angles*math.pi)
acosi = tf.stack((aco,asi),axis=-1)
x = tf.concat((x, acosi), axis=1)
x = tf.identity(x, name="input")
NCAP = 128
NMAPS = 128
SMAP = 7 # size of map
# we have different activations for different maps
x = tf.layers.dense(x, 1024, activation=tf.nn.relu)
cap_r = tf.layers.dense(x, NCAP * 2, activation=None)
cap_r = tf.reshape(cap_r, (-1, NCAP, 2))
# [-1,1] - > [0,SMAP-1] (to index space)
cap_r = (cap_r + 1) * (SMAP - 1) / 2.0
# all acap is positive because of sigmoid
cap_a = tf.layers.dense(x, NCAP * NMAPS, activation=tf.nn.sigmoid)
cap_a = tf.reshape(cap_a, (-1, NCAP, NMAPS))
x = caps.caps_tf(cap_r, cap_a, SMAP)
x = tf.layers.conv2d_transpose(x, 64, 4, 2, padding='same', activation=tf.nn.relu)
x = tf.layers.conv2d_transpose(x, 1, 4, 2, padding='same', activation=tf.nn.tanh)
x = tf.identity(x, name="output")
return x
def reconstructor_tf(x, angles, reuse=True):
with tf.variable_scope("Reconstructor", reuse=reuse):
aco = tf.cos(angles * math.pi)
asi = tf.sin(angles * math.pi)
acosi = tf.stack((aco, asi), axis=-1)
x = tf.concat((x, acosi), axis=1)
x = tf.identity(x, name="input")
CAP_LEN = 10
NCAP = 20
NCATEGORICAL = 10
NMAPS = 32
SMAP = 7 # size of map
x_categorical = tf.nn.softmax(x[:, :NCATEGORICAL])
x_pose = x[:, NCATEGORICAL:]
x_firstcaps = tf.layers.dense(x_pose, CAP_LEN * NCATEGORICAL * NCAP)
x_firstcaps = tf.reshape(x_firstcaps,
(-1, NCAP, CAP_LEN, NCATEGORICAL))
x_firstcaps = tf.einsum('bcln,bn->bcl', x_firstcaps, x_categorical)
cap_r = []
cap_a = []
for i in range(NCAP):
x = x_firstcaps[:, i, :]
x_r = tf.layers.dense(x, 128, activation=tf.nn.relu)
x_r = tf.layers.dense(x_r, 2, activation=None)
x_a = tf.layers.dense(x, 128, activation=tf.nn.relu)
x_a = tf.layers.dense(x_a, NMAPS, activation=tf.nn.sigmoid)
cap_r.append(x_r)
cap_a.append(x_a)
cap_r = tf.stack(cap_r, axis=1)
cap_a = tf.stack(cap_a, axis=1)
# [-1,1] - > [0,SMAP-1] (to index space)
cap_r = (cap_r + 1) * (SMAP - 1) / 2.0
x = caps.caps_tf(cap_r, cap_a, SMAP)
x = tf.layers.conv2d_transpose(x,
32,
4,
2,
padding='same',
activation=tf.nn.relu)
x = tf.layers.conv2d_transpose(x,
1,
4,
2,
padding='same',
activation=tf.nn.tanh)
x = tf.identity(x, name="output")
return x