Skip to content

Latest commit

 

History

History
123 lines (88 loc) · 8.18 KB

regressions-dl20-passage.bge-base-en-v1.5.hnsw-int8.onnx.md

File metadata and controls

123 lines (88 loc) · 8.18 KB
Raw

Anserini Regressions: TREC 2020 Deep Learning Track (Passage)

Model: BGE-base-en-v1.5 with quantized HNSW indexes (using ONNX for on-the-fly query encoding)

This page describes regression experiments, integrated into Anserini's regression testing framework, using the BGE-base-en-v1.5 model on the TREC 2020 Deep Learning Track passage ranking task, as described in the following paper:

Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597, 2023.

In these experiments, we are performing query inference "on-the-fly" with ONNX.

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.bge-base-en-v1.5.hnsw-int8.onnx

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.bge-base-en-v1.5.hnsw-int8.onnx

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-bge-base-en-v1.5.tar -P collections/
tar xvf collections/msmarco-passage-bge-base-en-v1.5.tar -C collections/

To confirm, msmarco-passage-bge-base-en-v1.5.tar is 59 GB and has MD5 checksum 353d2c9e72e858897ad479cca4ea0db1. 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.bge-base-en-v1.5.hnsw-int8.onnx \
  --corpus-path collections/msmarco-passage-bge-base-en-v1.5

Indexing

Sample indexing command, building quantized HNSW indexes:

bin/run.sh io.anserini.index.IndexHnswDenseVectors \
  -threads 16 \
  -collection JsonDenseVectorCollection \
  -input /path/to/msmarco-passage-bge-base-en-v1.5 \
  -generator DenseVectorDocumentGenerator \
  -index indexes/lucene-hnsw-int8.msmarco-v1-passage.bge-base-en-v1.5/ \
  -M 16 -efC 100 -quantize.int8 \
  >& logs/log.msmarco-passage-bge-base-en-v1.5 &

The path /path/to/msmarco-passage-bge-base-en-v1.5/ should point to the corpus downloaded above. Upon completion, we should have an index with 8,841,823 documents.

Note that here we are explicitly using Lucene's NoMergePolicy merge policy, which suppresses any merging of index segments. This is because merging index segments is a costly operation and not worthwhile given our query set. Furthermore, we are using Lucene's Automatic Byte Quantization feature, which increase the on-disk footprint of the indexes since we're storing both the int8 quantized vectors and the float32 vectors, but only the int8 quantized vectors need to be loaded into memory. See issue #2292 for some experiments reporting the performance impact.

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:

bin/run.sh io.anserini.search.SearchHnswDenseVectors \
  -index indexes/lucene-hnsw-int8.msmarco-v1-passage.bge-base-en-v1.5/ \
  -topics tools/topics-and-qrels/topics.dl20.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-passage-bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.dl20.txt \
  -encoder BgeBaseEn15 -hits 1000 -efSearch 1000 -threads 16 &

Evaluation can be performed using trec_eval:

bin/trec_eval -m map -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.dl20.txt
bin/trec_eval -m ndcg_cut.10 -c tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.dl20.txt
bin/trec_eval -m recall.100 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.dl20.txt
bin/trec_eval -m recall.1000 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.dl20.txt

Effectiveness

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

AP@1000 BGE-base-en-v1.5
DL20 (Passage) 0.465
nDCG@10 BGE-base-en-v1.5
DL20 (Passage) 0.678
R@100 BGE-base-en-v1.5
DL20 (Passage) 0.717
R@1000 BGE-base-en-v1.5
DL20 (Passage) 0.850

The above figures are from running brute-force search with cached queries on non-quantized flat indexes. With ONNX query encoding on quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more. Note that both HNSW indexing and quantization are non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking). 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.