matt-gardner
commented
8 years ago Use the create matrices operation, instead of train and test. Note that you'll also have to include "output": { "output matrices": true } at the same level as the "operation": { ... } parameters.
Alternatively, if you want to generate the features directly in scala/java code, instead of going through a file, you can just instantiate the feature generators, using something like this:
val params = new SpecFileReader("/dev/null").readSpecFile(specFile)
// Here we need to set up some stuff to use the SFE code. The first few things don't matter, but
// we'll use the graph and feature generators in the code below.
val praBase = "/dev/null"
val relation = "relation doesn't matter"
val outputter = Outputter.justLogger
val relationMetadata = new RelationMetadata(JNothing, praBase, outputter, fileUtil)
val graph = Graph.create(params \ "graph", praBase, outputter, fileUtil).get
val nodeFeatureGenerator = new NodeSubgraphFeatureGenerator(
params \ "node features",
relation,
relationMetadata,
outputter,
fileUtil
)
val nodePairFeatureGenerator = new NodePairSubgraphFeatureGenerator(
params \ "node pair features",
relation,
relationMetadata,
outputter,
fileUtil
)Then, to actually get features for a particular node pair, you do something like this:
println(s"Computing feature vector for entity pair ($mid1, $mid2)")
val instance = new NodePairInstance(graph.getNodeIndex(mid1), graph.getNodeIndex(mid2), true, graph)
val subgraph = nodePairFeatureGenerator.getLocalSubgraph(instance)
val features = nodePairFeatureGenerator.extractFeaturesAsStrings(instance, subgraph)You'd probably have to change a few things in that code, but that's the general idea.
samehkamaleldin
arthurcgusmao
Is there any documentation about how to use SFE as a library to extract SFE features into a file. without doing any learning or prediction.