def iv_net(x, training=True):
with tf.variable_scope('iv_net', reuse=tf.AUTO_REUSE):
x = ms.fc(x, units=40 * 40 * 32)
x = ms.bn(x, training=training)
x = tf.reshape(x, [-1, 40, 40, 32])
x = fn.id_blk(x, 32, [3, 3])
x = fn.id_blk(x, 32, [3, 3])
x = ms.max_pool(x, [2, 2], 2)
x = ms.conv(x, 64, [3, 3])
x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = fn.id_blk(x, 64, [3, 3])
x = fn.id_blk(x, 64, [3, 3])
x = ms.max_pool(x, [2, 2], 2)
x = ms.conv(x, 128, [3, 3])
x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = fn.id_blk(x, 128, [3, 3])
x = fn.id_blk(x, 128, [3, 3])
x = ms.max_pool(x, [2, 2], 2)
x = ms.conv(x, 256, [3, 3])
x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = fn.id_blk(x, 256, [3, 3])
x = fn.id_blk(x, 256, [3, 3])
x = tf.reshape(x, [-1, 5 * 5 * 256])
x = ms.fc(x, units=5)
x = ms.activation(x, relu=False)
return x
def fw_net(x, training=True):
with tf.name_scope('fw_net'):
with tf.variable_scope('fc1'):
x = ms.fc(x, 5 * 256)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
x = tf.reshape(x, [-1, 5, 256])
with tf.variable_scope('identical_conv1'):
x = ms.conv1d(x, 256, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 256, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 256, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 256, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
print(x.shape) # (128, 5, 256)
with tf.variable_scope('t_conv1'):
x = ms.t_conv1d(x, tf.Variable(tf.random.normal([3, 128, 256])),
(128, 10, 128), 2)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
with tf.variable_scope('identical_conv2'):
x = ms.conv1d(x, 128, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 128, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 128, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 128, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
with tf.variable_scope('t_conv2'):
x = ms.t_conv1d(x, tf.Variable(tf.random.normal((3, 64, 128))),
(128, 20, 64), 2)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
with tf.variable_scope('identical_conv3'):
x = ms.conv1d(x, 64, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 64, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 64, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 64, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
with tf.variable_scope('t_conv3'):
x = ms.t_conv1d(x, tf.Variable(tf.random.normal((3, 32, 64))),
(128, 40, 32), 2)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
with tf.variable_scope('identical_conv4'):
x = ms.conv1d(x, 32, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 32, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 32, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv1d(x, 32, 3)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = tf.reshape(x, [-1, 40 * 32])
with tf.variable_scope('fc2'):
x = ms.fc(x, units=603)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=False)
return x
def fw_net(x, training=True):
with tf.name_scope('fw_net'):
with tf.variable_scope('fw_net'):
with tf.name_scope('fc1'):
with tf.variable_scope('fc1'):
x = ms.fc(x, 5 * 5 * 256)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
x = tf.reshape(x, [-1, 5, 5, 256])
with tf.name_scope('identical_conv1'):
with tf.variable_scope('identical_conv1'):
x = ms.conv(x, 256, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 256, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 256, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 256, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 256, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
with tf.name_scope('t_conv1'):
with tf.variable_scope('t_conv1'):
x = ms.t_conv2d(x, 128, [2, 2], 2)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
with tf.name_scope('identical_conv2'):
with tf.variable_scope('identical_conv2'):
x = ms.conv(x, 128, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 128, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 128, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 128, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 128, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
with tf.name_scope('t_conv2'):
with tf.variable_scope('t_conv2'):
x = ms.t_conv2d(x, 64, [2, 2], 2)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
with tf.name_scope('identical_conv3'):
with tf.variable_scope('identical_conv3'):
x = ms.conv(x, 64, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 64, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 64, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 64, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 64, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
with tf.name_scope('t_conv3'):
with tf.variable_scope('t_conv3'):
x = ms.t_conv2d(x, 32, [2, 2], 2)
# x = ms.bn(x)
x = ms.activation(x, relu=True)
with tf.name_scope('identical_conv4'):
with tf.variable_scope('identical_conv4'):
x = ms.conv(x, 32, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 32, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 32, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 32, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = ms.conv(x, 32, [3, 3])
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=True)
x = tf.reshape(x, [-1, 40 * 40 * 32])
with tf.name_scope('fc2'):
with tf.variable_scope('fc2'):
x = ms.fc(x, units=603)
# x = ms.bn(x, training=training)
x = ms.activation(x, relu=False)
return x
t = ms.conv(t, 32, [3, 3])
# x = ms.bn(x, training=training)
t = ms.activation(t, relu=True)
t = ms.conv(t, 32, [3, 3])
# x = ms.bn(x, training=training)
t = ms.activation(t, relu=True)
t = ms.conv(t, 32, [3, 3])
# x = ms.bn(x, training=training)
t = ms.activation(t, relu=True)
t = tf.reshape(t, [-1, 40 * 40 * 32])
with tf.variable_scope('fc2'):
t = ms.fc(t, units=402)
# x = ms.bn(x, training=training)
t = ms.activation(t, relu=False)
new_loss = ms.huber_loss(t, y_)
new_optimizer = tf.train.AdamOptimizer(learning_rate=LR,
beta1=0.9,
beta2=0.999,
epsilon=1e-8,
name='new_adam')
new_op = new_optimizer.minimize(loss=new_loss)
var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='trans_part')