Skip to content

Latest commit

 

History

History
114 lines (81 loc) · 7.29 KB

regressions-dl20-passage-cos-dpr-distil-fw.md

File metadata and controls

114 lines (81 loc) · 7.29 KB
Raw

Anserini Regressions: TREC 2020 Deep Learning Track (Passage)

Model: cosDPR-distil with inverted indexes using the "fake-words" technique (q=40); pre-encoded queries

This page describes regression experiments, integrated into Anserini's regression testing framework, using the cosDPR-distil model on the TREC 2020 Deep Learning Track passage ranking task. In these experiments, we are using pre-encoded queries (i.e., cached results of query encoding).

Note that the NIST relevance judgments provide far more relevant passages per topic, unlike the "sparse" judgments provided by Microsoft (these are sometimes called "dense" judgments to emphasize this contrast). For additional instructions on working with MS MARCO passage collection, refer to this page.

The exact configurations for these regressions are stored in this YAML file. Note that this page is automatically generated from this template as part of Anserini's regression pipeline, so do not modify this page directly; modify the template instead and then run bin/build.sh to rebuild the documentation.

From one of our Waterloo servers (e.g., orca), the following command will perform the complete regression, end to end:

python src/main/python/run_regression.py --index --verify --search --regression dl20-passage-cos-dpr-distil-fw

We make available a version of the MS MARCO Passage Corpus that has already been encoded with cosDPR-distil.

From any machine, the following command will download the corpus and perform the complete regression, end to end:

python src/main/python/run_regression.py --download --index --verify --search --regression dl20-passage-cos-dpr-distil-fw

The run_regression.py script automates the following steps, but if you want to perform each step manually, simply copy/paste from the commands below and you'll obtain the same regression results.

Corpus Download

Download the corpus and unpack into collections/:

wget https://rgw.cs.uwaterloo.ca/pyserini/data/msmarco-passage-cos-dpr-distil.tar -P collections/
tar xvf collections/msmarco-passage-cos-dpr-distil.tar -C collections/

To confirm, msmarco-passage-cos-dpr-distil.tar is 57 GB and has MD5 checksum e20ffbc8b5e7f760af31298aefeaebbd. With the corpus downloaded, the following command will perform the remaining steps below:

python src/main/python/run_regression.py --index --verify --search --regression dl20-passage-cos-dpr-distil-fw \
  --corpus-path collections/msmarco-passage-cos-dpr-distil

Indexing

Sample indexing command, applying inverted indexes to dense vectors using the "fake-words" technique:

target/appassembler/bin/IndexInvertedDenseVectors \
  -collection JsonDenseVectorCollection \
  -input /path/to/msmarco-passage-cos-dpr-distil \
  -generator InvertedDenseVectorDocumentGenerator \
  -index indexes/lucene-index.msmarco-passage-cos-dpr-distil.fw-40/ \
  -threads 16 -encoding fw -fw.q 40 \
  >& logs/log.msmarco-passage-cos-dpr-distil &

The path /path/to/msmarco-passage-cos-dpr-distil/ should point to the corpus downloaded above.

Upon completion, we should have an index with 8,841,823 documents.

Retrieval

Topics and qrels are stored here, which is linked to the Anserini repo as a submodule. The regression experiments here evaluate on the 54 topics for which NIST has provided judgments as part of the TREC 2020 Deep Learning Track. The original data can be found here.

After indexing has completed, you should be able to perform retrieval as follows:

target/appassembler/bin/SearchInvertedDenseVectors \
  -index indexes/lucene-index.msmarco-passage-cos-dpr-distil.fw-40/ \
  -topics tools/topics-and-qrels/topics.dl20.cos-dpr-distil.jsonl.gz \
  -topicReader JsonIntVector \
  -output runs/run.msmarco-passage-cos-dpr-distil.cos-dpr-distil-fw-40.topics.dl20.cos-dpr-distil.jsonl.txt \
  -topicField vector -threads 16 -encoding fw -fw.q 40 -hits 1000 &

Evaluation can be performed using trec_eval:

target/appassembler/bin/trec_eval -m map -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-cos-dpr-distil.cos-dpr-distil-fw-40.topics.dl20.cos-dpr-distil.jsonl.txt
target/appassembler/bin/trec_eval -m ndcg_cut.10 -c tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-cos-dpr-distil.cos-dpr-distil-fw-40.topics.dl20.cos-dpr-distil.jsonl.txt
target/appassembler/bin/trec_eval -m recall.100 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-cos-dpr-distil.cos-dpr-distil-fw-40.topics.dl20.cos-dpr-distil.jsonl.txt
target/appassembler/bin/trec_eval -m recall.1000 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-cos-dpr-distil.cos-dpr-distil-fw-40.topics.dl20.cos-dpr-distil.jsonl.txt

Effectiveness

With the above commands, you should be able to reproduce the following results:

AP@1000 cosDPR-distill
DL20 (Passage) 0.4597
nDCG@10 cosDPR-distill
DL20 (Passage) 0.6666
R@100 cosDPR-distill
DL20 (Passage) 0.6909
R@1000 cosDPR-distill
DL20 (Passage) 0.8194

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run. Nevertheless, scores are generally within 0.005 of the reference values recorded in our YAML configuration file.

Also note that retrieval metrics are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking). Also, for computing nDCG, remember that we keep qrels of all relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the -l 2 option in trec_eval). The experimental results reported here are directly comparable to the results reported in the track overview paper.

Reproduction Log*

To add to this reproduction log, modify this template and run bin/build.sh to rebuild the documentation.