Sharing weights in Tensorflow between two subgraphs

Question

I have the following setup, where each input consists of two trajectories. I want that the left graph has the same weight as the right graph

I tried to follow the approach described here for sharing variables, https://www.tensorflow.org/versions/r1.0/how_tos/variable_scope/, however it is not working. Two different graphs are created. What am I doing wrong?

def build_t_model(trajectories):
    """
    Function to build a subgraph
    """
    with tf.name_scope('h1_t'):
        weights = tf.Variable(tf.truncated_normal([150, h1_t_units], stddev=1.0/math.sqrt(float(150))), name='weights')
        biases = tf.Variable(tf.zeros([h1_t_units]), name='biases')
        h1_t = tf.nn.relu(tf.matmul(trajectories, weights) + biases)

    with tf.name_scope('h2_t'):
        weights = tf.Variable(tf.truncated_normal([h1_t_units, h2_t_units], stddev=1.0/math.sqrt(float(h1_t_units))), name='weights')
        biases = tf.Variable(tf.zeros([h2_t_units]), name='biases')
        h2_t = tf.nn.relu(tf.matmul(h1_t, weights) + biases)

    with tf.name_scope('h3_t'):
        weights = tf.Variable(tf.truncated_normal([h2_t_units, M], stddev=1.0/math.sqrt(float(h2_t_units))), name='weights')
        biases = tf.Variable(tf.zeros([M]), name='biases')
        h3_t = tf.nn.relu(tf.matmul(h2_t, weights) + biases)

    return h3_t


# We build two trajectory networks. The weights should be shared
with tf.variable_scope('traj_embedding') as scope:        
    self.embeddings_left = build_t_model(self.input_traj)
    scope.reuse_variables()
    self.embeddings_right = build_t_model(self.input_traj_mv)

Answer 1

Okay, use tf.get_variable instead of tf.Variable for this. This works

def build_t_model(trajectories):
            """
            Build the trajectory network
            """
            with tf.name_scope('h1_t'):
                weights = tf.get_variable(
                    'weights1', 
                    shape=[150, h1_t_units],
                    initializer=tf.truncated_normal_initializer(
                        stddev=1.0/math.sqrt(float(150))))
                biases = tf.get_variable(
                    'biases1', 
                    initializer=tf.zeros_initializer(shape=[h1_t_units]))
                h1_t = tf.nn.relu(tf.matmul(trajectories, weights) + biases)

            with tf.name_scope('h2_t'):
                weights = tf.get_variable(
                    'weights2', 
                    shape=[h1_t_units, h2_t_units],
                    initializer=tf.truncated_normal_initializer(
                        stddev=1.0/math.sqrt(float(h1_t_units))))
                biases = tf.get_variable(
                    'biases2', 
                    initializer=tf.zeros_initializer(shape=[h2_t_units]))
                h2_t = tf.nn.relu(tf.matmul(h1_t, weights) + biases)

            with tf.name_scope('h3_t'):
                weights = tf.get_variable(
                    'weights3', 
                    shape=[h2_t_units, M],
                    initializer=tf.truncated_normal_initializer(
                        stddev=1.0/math.sqrt(float(h2_t_units))))
                biases = tf.get_variable(
                    'biases3', 
                    initializer=tf.zeros_initializer(shape=[M]))
                h3_t = tf.nn.relu(tf.matmul(h2_t, weights) + biases)
            return h3_t