nn.py 文件源码

def double_MLP(self, inputs, n_splits=1):
    """"""

    batch_size = tf.shape(inputs)[0]
    bucket_size = tf.shape(inputs)[1]
    input_size = inputs.get_shape().as_list()[-1]
    output_size = self.mlp_size
    output_shape = tf.pack([batch_size, bucket_size, bucket_size, output_size])
    shape_to_set = [tf.Dimension(None), tf.Dimension(None), tf.Dimension(None), tf.Dimension(output_size)]

    if self.moving_params is None:
      if self.drop_gradually:
        s = self.global_sigmoid
        keep_prob = s + (1-s)*self.mlp_keep_prob
      else:
        keep_prob = self.mlp_keep_prob
    else:
      keep_prob = 1
    if isinstance(keep_prob, tf.Tensor) or keep_prob < 1:
      noise_shape = tf.pack([batch_size, 1, input_size])
      inputs = tf.nn.dropout(inputs, keep_prob, noise_shape=noise_shape)

    lin1, lin2 = linalg.linear(inputs,
                               output_size*n_splits,
                               n_splits=2,
                               add_bias=True,
                               moving_params=self.moving_params)
    lin1 = tf.reshape(tf.transpose(lin1, [0, 2, 1]), tf.pack([-1, bucket_size, 1]))
    lin2 = tf.reshape(tf.transpose(lin2, [0, 2, 1]), tf.pack([-1, 1, bucket_size]))
    lin = lin1 + lin2
    lin = tf.reshape(lin, tf.pack([batch_size, n_splits*output_size, bucket_size, bucket_size]))
    lin = tf.transpose(lin, [0,2,3,1])
    top_mlps = tf.split(3, n_splits, self.mlp_func(lin))
    for top_mlp in top_mlps:
      top_mlp.set_shape(shape_to_set)
    if self.moving_params is None:
      with tf.variable_scope('Linear', reuse=True):
        matrix = tf.get_variable('Weights')
        I = tf.diag(tf.ones([self.mlp_size]))
        for W in tf.split(1, 2*n_splits, matrix):
          WTWmI = tf.matmul(W, W, transpose_a=True) - I
          tf.add_to_collection('ortho_losses', tf.nn.l2_loss(WTWmI))
      for split in top_mlps:
        tf.add_to_collection('covar_losses', self.covar_loss(split))
    if n_splits == 1:
      return top_mlps[0]
    else:
      return top_mlps

  #=============================================================