GraphSAGE 代码解析(三) - aggregators.py
1. class MeanAggregator(Layer):该类主要用于实现 1. __init__()?__init_() 用于获取并初始化成员变量 dropout,bias(False),act(ReLu),concat(False),input_dim,output_dim,name(Variable scopr) 用glorot()方法初始化节点v的权值矩阵 vars[‘self_weights‘] 和邻居节点均值u的权值矩阵 vars[‘neigh_weights‘] 用零向量初始化vars[‘bias‘]。(见inits.py: zeros(shape)) 若logging为True,则调用 layers.py 中 class Layer()的成员函数_log_vars(),生成vars中各个变量的直方图。 glorot() 其中,glorot()?在inits.py中定义,用于权值初始化。(from .inits?import glorot) 均匀分布初始化方法,又称Xavier均匀初始化,参数从 [-limit,limit] 的均匀分布产生,其中limit为 sqrt(6 / (fan_in + fan_out))。fan_in为权值张量的输入单元数,fan_out是权重张量的输出单元数。该函数返回 [fan_in,fan_out]大小的Variable。 1 def glorot(shape,name=None): 2 """Glorot & Bengio (AISTATS 2010) init.""" 3 init_range = np.sqrt(6.0/(shape[0]+shape[1])) 4 initial = tf.random_uniform(shape,minval=-init_range,maxval=init_range,dtype=tf.float32) 5 return tf.Variable(initial,name=name) 2. _call(inputs)class MeanAggregator(Layer) 中的 _call(inputs) 函数是对父类class Layer(object)方法_call(inputs)的重写。 用于实现最上方的迭代更新式子。 在layer.py 中定义的?class Layer(object)中,执行特殊函数def __call__(inputs) 时有: outputs = self._call(inputs)调用_call(inputs) 方法,也即在这里调用子类MeanAggregator(Layer)中的_call(inputs)方法。 tf.nn.dropout(x,keep_prob,noise_shape=None,seed=None,name=None) With probability keep_prob,outputs the input element scaled up by 1 / keep_prob,otherwise outputs 0. The scaling is so that the expected sum is unchanged. 注意:输出的非0元素是原来的 “1/keep_prob” 倍,以保证总和不变。 tf.add_n(inputs,name=None) Adds all input tensors element-wise. Args: inputs: A list of Tensor or IndexedSlices objects,each with same shape and type. name: A name for the operation (optional). Returns: A Tensor of same shape and type as the elements of inputs. Raises: ValueError: If inputs don‘t all have same shape and dtype or the shape cannot be inferred. output = tf.concat([from_self,from_neighs],axis=1) 这里注意在concat后其维数变为之前的2倍。 3. class MeanAggregator(Layer) 代码1 class MeanAggregator(Layer): 2 """ 3 Aggregates via mean followed by matmul and non-linearity. 4 """ 5 6 def __init__(self,neigh_input_dim=None, 7 dropout=0.,bias=False,act=tf.nn.relu, 8 name=None,concat=False,**kwargs): 9 super(MeanAggregator,self).__init__(**kwargs) 10 11 self.dropout = dropout 12 self.bias = bias 13 self.act = act 14 self.concat = concat 15 16 if neigh_input_dim is None: 17 neigh_input_dim = input_dim 18 19 if name is not None: 20 name = ‘/‘ + name 21 else: 22 name = ‘‘ 23 24 with tf.variable_scope(self.name + name + ‘_vars‘): 25 self.vars[‘neigh_weights‘] = glorot([neigh_input_dim,output_dim],26 name=‘neigh_weights‘) 27 self.vars[‘self_weights‘] = glorot([input_dim,28 name=‘self_weights‘) 29 if self.bias: 30 self.vars[‘bias‘] = zeros([self.output_dim],name=‘bias‘) 31 32 if self.logging: 33 self._log_vars() 34 35 self.input_dim = input_dim 36 self.output_dim = output_dim 37 38 def _call(self,inputs): 39 self_vecs,neigh_vecs = inputs 40 41 neigh_vecs = tf.nn.dropout(neigh_vecs,1-self.dropout) 42 self_vecs = tf.nn.dropout(self_vecs,1-self.dropout) 43 neigh_means = tf.reduce_mean(neigh_vecs,axis=1) 44 45 # [nodes] x [out_dim] 46 from_neighs = tf.matmul(neigh_means,self.vars[‘neigh_weights‘]) 47 48 from_self = tf.matmul(self_vecs,self.vars["self_weights"]) 49 50 if not self.concat: 51 output = tf.add_n([from_self,from_neighs]) 52 else: 53 output = tf.concat([from_self,axis=1) 54 55 # bias 56 if self.bias: 57 output += self.vars[‘bias‘] 58 59 return self.act(output) 2. class GCNAggregator(Layer)这里__init__()与MeanAggregator基本相同,在_call()的实现中略有不同。 1 def _call(self,inputs): 2 self_vecs,neigh_vecs = inputs 3 4 neigh_vecs = tf.nn.dropout(neigh_vecs,1-self.dropout) 5 self_vecs = tf.nn.dropout(self_vecs,1-self.dropout) 6 means = tf.reduce_mean(tf.concat([neigh_vecs, 7 tf.expand_dims(self_vecs,axis=1)],axis=1),axis=1) 8 9 # [nodes] x [out_dim] 10 output = tf.matmul(means,self.vars[‘weights‘]) 11 12 # bias 13 if self.bias: 14 output += self.vars[‘bias‘] 15 16 return self.act(output) 其中对means求解时, 1. 先将self_vecs行列转换(tf.expand_dims(self_vecs,axis=1)), 2. 之后self_vecs的行数与neigh_vecs行数相同时,将二者concat,即相当于在原先的neigh_vecs矩阵后面新增一列self_vecs的转置 3. 最后将得到的矩阵每行求均值,即得means. 之后means与权值矩阵vars[‘weights‘]求内积,并加上vars[‘bias‘],最终将该值带入激活函数(ReLu)。 下面举个例子简单说明(例子中省略了点乘W的操作): 1 import tensorflow as tf 2 3 neigh_vecs = [[1,2,3],[4,5,6],[7,8,9]] 4 self_vecs = [2,3,4] 5 6 means = tf.reduce_mean(tf.concat([neigh_vecs, 7 tf.expand_dims(self_vecs,axis=1) 8 9 print(tf.shape(self_vecs)) 10 11 print(tf.expand_dims(self_vecs,axis=0)) 12 # Tensor("ExpandDims_1:0",shape=(1,3),dtype=int32) 13 14 print(tf.expand_dims(self_vecs,axis=1)) 15 # Tensor("ExpandDims_2:0",shape=(3,1),dtype=int32) 16 17 sess = tf.Session() 18 print(sess.run(tf.expand_dims(self_vecs,axis=1))) 19 # [[2] 20 # [3] 21 # [4]] 22 23 print(sess.run(tf.concat([neigh_vecs,24 tf.expand_dims(self_vecs,axis=1))) 25 # [[1 2 3 2] 26 # [4 5 6 3] 27 # [7 8 9 4]] 28 29 print(means) 30 # Tensor("Mean:0",),dtype=int32) 31 32 print(sess.run(tf.reduce_mean(tf.concat([neigh_vecs,33 tf.expand_dims(self_vecs,axis=1))) 34 # [2 4 7] 35 36 # [[1 2 3 2] = 8 // 4 = 2 37 # [4 5 6 3] = 18 // 4 = 4 38 # [7 8 9 4]] = 28 // 4 = 7 39 40 bias = [1] 41 output = means + bias 42 print(sess.run(output)) 43 # [3 5 8] 44 # [2 + 1,4 + 1,7 + 1] = [3,8] (编辑:李大同) 【声明】本站内容均来自网络,其相关言论仅代表作者个人观点,不代表本站立场。若无意侵犯到您的权利,请及时与联系站长删除相关内容! |