CODEWITHSUNDEEP
pythonpython-3.xtensorflow
shakedzy
shakedzy

Reputation: 2893

Tensorflow: unable to share dense/kernel - ValueError: Trying to share variable dense/kernel, but specified shape (100, 160) and found shape (9, 100)

I built a Deep-Q-Network using Tensorflow. When I try to create two of those (I would like to have the network playing against itself), I get:

ValueError: Trying to share variable dense/kernel, but specified shape (100, 160) and found shape (9, 100).

This is my network:

class QNetwork:
    """
    A Q-Network implementation
    """
    def __init__(self, input_size, output_size, hidden_layers_size, gamma, maximize_entropy, reuse):
        self.q_target = tf.placeholder(shape=(None, output_size), dtype=tf.float32)
        self.r = tf.placeholder(shape=None, dtype=tf.float32)
        self.states = tf.placeholder(shape=(None, input_size), dtype=tf.float32)
        self.enumerated_actions = tf.placeholder(shape=(None, 2), dtype=tf.int32)
        self.learning_rate = tf.placeholder(shape=[], dtype=tf.float32)
        layer = self.states
        for l in hidden_layers_size:
            layer = tf.layers.dense(inputs=layer, units=l, activation=tf.nn.relu,
                                    kernel_initializer=tf.contrib.layers.xavier_initializer(),
                                    reuse=reuse)
        self.output = tf.layers.dense(inputs=layer, units=output_size,
                                      kernel_initializer=tf.contrib.layers.xavier_initializer(),
                                      reuse=reuse)
        self.predictions = tf.gather_nd(self.output, indices=self.enumerated_actions)
        if maximize_entropy:
            self.future_q = tf.log(tf.reduce_sum(tf.exp(self.q_target), axis=1))
        else:
            self.future_q = tf.reduce_max(self.q_target, axis=1)
        self.labels = self.r + (gamma * self.future_q)
        self.cost = tf.reduce_mean(tf.losses.mean_squared_error(labels=self.labels, predictions=self.predictions))
        self.optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(self.cost)

And this code fails:

q1 = QNetwork(9, 9, [100, 160, 160, 100], gamma=0.99, maximize_entropy=False, reuse=tf.AUTO_REUSE)
q2 = QNetwork(9, 9, [100, 160, 160, 100], gamma=0.99, maximize_entropy=False, reuse=tf.AUTO_REUSE)

Any idea how to solve this? (Running TF 1.10.1, Python 3.6.5)

Upvotes: 2

Views: 1735

Answers (1)

shakedzy
shakedzy

Reputation: 2893

Solved.

I needed to:

  • Give each layer a unique name
  • Put everything in a variable_scope with reuse=tf.AUTO_REUSE (for the Adam optimizer)

Upvotes: 3

Related Questions