For reference of using tf.get_variable

conv_tf.py

def conv2d(inputs, filter, strides, name='conv2d'):
        k = tf.get_variable('W', filter, initializer=xavier_initializer_conv2d())
        b = tf.get_variable('b', filter[-1], initializer=tf.constant_initializer(0.0))
        conv = tf.nn.conv2d(inputs, k, strides, 'SAME')
        bias_add = tf.nn.bias_add(conv, b)
        return tf.nn.relu(bias_add, name=name)


def vision_model(frames, n_frames):
    with tf.variable_scope('Conv1') as scope:
        conv1 = conv2d(frames, [8, 8, n_frames, 32], [1, 4, 4, 1], scope.name)

    with tf.variable_scope('Conv2') as scope:
        conv2 = conv2d(conv1, [4, 4, 32, 64], [1, 2, 2, 1], scope.name)

    with tf.variable_scope('Conv3') as scope:
        conv3 = conv2d(conv2, [3, 3, 64, 64], [1, 1, 1, 1], scope.name)

    return conv3


def atari_cnn(inputs, batch_size, n_actions, n_frames, name='atari_cnn'):
    """
    input 84 x 84 x 4 (image is 84x84, grayscaled)
    32 filters 8 x 8 stride 4, relu
    64 filters 4 x 4 stride 2, relu
    64 filters 3 x 3 stride 1, relu
    fc 512 units, relu
    fc 4-18 units (actions)
    """
    with tf.variable_scope(name):

        conv3 = vision_model(inputs, n_frames)

        reshaped = tf.reshape(conv3, [batch_size, -1])
        dim = reshaped.get_shape()[1].value

        with tf.variable_scope('Dense1') as scope:
            w = tf.get_variable('W', [dim, 512], initializer=xavier_initializer())
            b = tf.get_variable('b', [512], initializer=tf.constant_initializer(0.0))
            dense1 = tf.nn.relu(tf.matmul(reshaped, w) + b, name=scope.name)

        with tf.variable_scope('Dense2') as scope:
            w = tf.get_variable('W', [512, n_actions], initializer=xavier_initializer())
            b = tf.get_variable('b', [n_actions], initializer=tf.constant_initializer(0.0))
            dense2 = tf.add(tf.matmul(dense1, w), b, name=scope.name)

        return dense2