simple.py 文件源码

def dense(x,
          num_outputs,
          scope=None,
          activation=None,
          reuse=None,
          bn=False,
          post_bn=False,
          phase=None):

    with tf.variable_scope(scope, 'dense', reuse=reuse):
        # convert x to 2-D tensor
        dim = np.prod(x._shape_as_list()[1:])
        x = tf.reshape(x, [-1, dim])
        weights_shape = (x.get_shape().dims[-1], num_outputs)

        # dense layer
        weights = tf.get_variable('weights', weights_shape,
                                  initializer=variance_scaling_initializer())
        biases = tf.get_variable('biases', [num_outputs],
                                 initializer=tf.zeros_initializer)
        output = tf.matmul(x, weights) + biases
        if bn: output = batch_norm(output, phase, scope='bn')
        if activation: output = activation(output)
        if post_bn: output = batch_norm(output, phase, scope='post_bn')

    return output