Skip to content

Latest commit

 

History

History
111 lines (87 loc) · 3.96 KB

inmemory_sampler.md

File metadata and controls

111 lines (87 loc) · 3.96 KB
Raw

The TF-GNN In Memory Sampler

Overview

In many applications graphs are small enough to fit on a single worker machine memory (e.g. OGB Node Property Prediction). The good representation for those small (heterogeneous) graphs is tfgnn.GraphTensor itself. The TF-GNN library provides set of tools that allow graph sampling from those small graphs using graph schema and sampling spec, as covered in the data prep guide.

Below we describe in-memory sampling using a toy example graph. For complete end-to-end story see Solving OGBN-MAG end-to-end.

Sampling from GraphTensor A,B,C

For the sake of example, let's consider toy homogeneous graph with nodes A, B, C and edges A->{B, C}, B->{A}, C->{B}:

import tensorflow as tf
import tensorflow_gnn as tfgnn

full_graph = tfgnn.homogeneous(
    tf.constant([0, 0, 1, 2]),
    tf.constant([1, 2, 0, 1]),
    node_features={'name': tf.constant(['A', 'B', 'C'])},
)

NOTE: for real applications, the full graph could be stored as tf.Example, e.g. using tfgnn.write_example, and loaded using tfgnn.parse_single_example.

Now, consider uniform sampling for two hops with sampling sizes [4, 2]. The library provides tfgnn.sampler.SamplingSpecBuilder() class that simplifies sampling spec construction:

graph_schema = tfgnn.create_schema_pb_from_graph_spec(full_graph)

sampling_spec_builder = tfgnn.sampler.SamplingSpecBuilder(
    graph_schema,
    default_strategy=tfgnn.sampler.SamplingStrategy.RANDOM_UNIFORM)

seed = sampling_spec_builder.seed("nodes")
hop1 = seed.sample(4, "edges")
hop2 = hop1.sample(2, "edges")

sampling_spec = sampling_spec_builder.build()

The sampling could be encapsulated as a "sampling model", instance of tf.keras.Model, which takes as an input collection of seed node ids and returns sampled subgaphs as tfgnn.GraphTensor.

The sampler library provides helper function that constructs sampling models using graph schema, sampling spec and two factory methods: one for edges and one for node features (see Sampling API for mode details).

from tensorflow_gnn.experimental import sampler

def edge_sampler(sampling_op: tfgnn.sampler.SamplingOp):
  edge_set_name = sampling_op.edge_set_name
  sample_size = sampling_op.sample_size
  return sampler.InMemUniformEdgesSampler.from_graph_tensor(
      full_graph, edge_set_name, sample_size=sample_size, name=sampling_op.op_name
  )

def get_features(node_set_name: tfgnn.NodeSetName):
  return sampler.InMemIndexToFeaturesAccessor.from_graph_tensor(
      full_graph, node_set_name
  )

sampling_model = sampler.create_sampling_model_from_spec(
    graph_schema, sampling_spec, edge_sampler, get_features,
    seed_node_dtype=tf.int32)

Here the sampling_model takes indices of seed nodes in the full_graph as its input and returns sampled tfgnn.GraphTensor as result. The seeds are ragged rank 1 tensors with a shape [batch_size, (num_seeds)]. The 1st dimension is a batch dimension. It is used to create batches of sampled subgrahs. The 2nd ragged dimension allows to specify potentially multiple seed nodes to use a a starting points for sampling for each subgraph, as:

[[subgraph1/seed1, subgraph1/seed2,...], [subgraph2/seed1, ...], ...]

The result tfgnn.GraphTensor has shape [batch_size] with one graph component per graph.

The sampling model can be used with tf.data.Dataset to construct input data pipeline:

seeds = tf.data.Dataset.from_tensor_slices([0, 1, 2, 0])
# Create batches of up to two seeds
seeds = seeds.batch(2)
# [seed1, seed2] -> [[seed1], [seed2]]
seeds = seeds.map(
    lambda s: tf.RaggedTensor.from_row_lengths(s, tf.ones_like(s))
)
graphs = seeds.map(sampling_model)
for graph in graphs:
  print(graph)