31 lines
840 B
Python
31 lines
840 B
Python
|
from __future__ import division
|
||
|
|
from __future__ import print_function
|
||
|
|
|
||
|
|
from graphsage.layers import Layer
|
||
|
|
|
||
|
|
import tensorflow as tf
|
||
|
|
flags = tf.app.flags
|
||
|
|
FLAGS = flags.FLAGS
|
||
|
|
|
||
|
|
|
||
|
|
"""
|
||
|
|
Classes that are used to sample node neighborhoods during
|
||
|
|
convolutions.
|
||
|
|
"""
|
||
|
|
|
||
|
|
class UniformNeighborSampler(Layer):
|
||
|
|
"""
|
||
|
|
Uniformly samples neighbors.
|
||
|
|
Assumes that adj lists are padded with random re-sampling
|
||
|
|
"""
|
||
|
|
def __init__(self, adj_info, **kwargs):
|
||
|
|
super(UniformNeighborSampler, self).__init__(**kwargs)
|
||
|
|
self.adj_info = adj_info
|
||
|
|
|
||
|
|
def _call(self, inputs):
|
||
|
|
ids, num_samples = inputs
|
||
|
|
adj_lists = tf.nn.embedding_lookup(self.adj_info, ids)
|
||
|
|
adj_lists = tf.transpose(tf.random_shuffle(tf.transpose(adj_lists)))
|
||
|
|
adj_lists = tf.slice(adj_lists, [0,0], [-1, num_samples])
|
||
|
|
return adj_lists
|