Adding multi-threads support with multiple params to SearchCollection #470
Conversation
… list of params and construct the Lucene searcher in a for loop and output to different files. This can reduce the effor of reading index from disk every time one'd like to run another set of params. The reranking is also supported in the similar way
|
|
||
| import org.apache.lucene.search.similarities.Similarity; | ||
|
|
||
| public class AuxSimilarity { |
There was a problem hiding this comment.
How about we name this class TaggedSimilarity?
| package io.anserini.search.similarity; | ||
|
|
||
| import org.apache.lucene.search.similarities.Similarity; | ||
|
|
| searcher.close(); | ||
| final long durationMillis = TimeUnit.MILLISECONDS.convert(System.nanoTime() - start, TimeUnit.NANOSECONDS); | ||
| LOG.info("Total " + numTopics + " topics searched in " | ||
| + DurationFormatUtils.formatDuration(durationMillis, "HH:mm:ss")); | ||
| LOG.info("Total run time: " + DurationFormatUtils.formatDuration(durationMillis, "HH:mm:ss")); |
There was a problem hiding this comment.
Why don't we print out number of topics anymore?
There was a problem hiding this comment.
Because now it is print inside each retrieval thread.
Search for LOG.info("Run " + topics.size() in this PR
|
|
||
| private SearcherThread(IndexSearcher searcher, SortedMap<K, Map<String, String>> topics, AuxSimilarity auxSimilarity, | ||
| String cascadeTag, RerankerCascade cascade, String outputPath, String runTag) throws IOException { | ||
| this.searcher = searcher; |
There was a problem hiding this comment.
Why do we need a separate cascade tag? Each reranker has its own tag, right? So can't the RerankerCascade reconstruct the tag by joining each individual tag?
There was a problem hiding this comment.
Most of the time we will probably have just 1 reranker and I think it is better to have just 1 tag.
Also, we have TieBreaker as the default last reranker and I am not sure if it is worthy to concatenate all tags.
|
High-level comments, for discussion:
|
|
|
Can you create a separate issue for above so we don't lose track of it? Also, please make sure regressions pass before you merge? |
|
ALL Regression tests passed on tuna, going to merge |
token test bug fix
In this PR I fundamentally change how
SearchCollectionworks.-b 0.2 0.75for BM25 or-mu 200 2000for QL. One can also provide multiple params for reranking models, at the same time, for example-rm3.fbDocs 10 20 -rm3.fbTerms 50 100for RM3. As a result, there are N1N2N3*... run files being generated where N1..Nn are the values of the params.SearchCollectionspawns new threads for all retrievals by default. This, together with the newly introduced-inmemoption, make a bunch of retrieval less expensive since multiple retrievals need to load the index (potentially in the memory for better multi-threading) once.