尝试在Tensorflow中初始化神经网络，收到TypeError

Question

I am having troubles initializing my neural network properly in Tensorflow.

In my BayesianNN class, I have a build_graph function:

def build_graph():
 self._create_feedforward() 
 self._initializer()
 self._define_layers()
 self._regularization()

As of right now, my _create_feedforward() sets the framework for how I want my weights and biases initialized, as well as their output:

def _create_feedforward(self, input, output, scope):
    with tf.variable_scope(scope, reuse=tf.AUTO_REUSE) as scope:
        self.weights = tf.get_variable('weights',
                                       shape=(input.shape[1], output),
                                       initializer=tf.random_normal_initializer(0,0.01),
                                       dtype=tf.float32)
        self.biases = tf.get_variable('biases',
                                      shape=(output),
                                      initializer=tf.constant_initializer(0.0),
                                      dtype=tf.float32)

        activation = tf.matmul(input, self.weights) + self.biases

        return tf.matmul(tf.diag(self.Bern_prob.sample((input.shape[1],))), 
                                  activation) 

While my _define_layers function sets the input parameters for my network:

def _define_layers(self):
    layer_1_output = _create_feedforward(model_X, self.layer_1_dim, 'layer_1')
    layer_2_output = _create_feedforward(self.layer_1_dim, self.layer_2_dim, 'layer_2')
    layer_3_output = _create_feedforward(self.layer_2_dim, [1], 'layer_3')

Because _create_feedforward is called before _define_layers, I am receiving a typeError for not enough inputs. But I cant call _define_layers first because _create_feedforward has not been defined yet.

TypeError: _create_feedforward() missing 3 required positional arguments: 'input', 'output', and 'scope'

I know why this error is happening, but how about do I implement this in code neatly and without errors?

Answer 1

The issue here is that in _create_feedforward you have the argument self . However, when you call it in _define_layers , you don't call the function as part of a class. Try this:

def _define_layers(self):
    layer_1_output = self._create_feedforward(model_X, self.layer_1_dim, 'layer_1')
    layer_2_output = self._create_feedforward(self.layer_1_dim, self.layer_2_dim,'layer_2')
    layer_3_output = self._create_feedforward(self.layer_2_dim, [1], 'layer_3')