stanis-morozov / rpg

A supplementary code for Relevance Proximity Graphs for Fast Relevance Retrieval

Relevance Proximity Graphs

A supplementary code for Relevance Proximity Graphs for Fast Relevance Retrieval. This implementation is substantially based on HNSW code https://github.com/nmslib/hnswlib.

What does it do?

It performs a fast relevance retrieval over a large-scale database with a given relevance function. The relevance function is defined on the (query, item) pairs without predefined similarity measure between two items or two queries.

What do I need to run it?

• Use any popular 64-bit Linux operating system
  • Tested on Ubuntu16.04, should work fine on most linux x64 and even MacOS;
• Run the Makefile to compile c++ sources
  • sudo apt-get install --upgrade gcc g++ libstdc++6 make
  • make
  • chmod u+x download.py compute_scores.py
• To work with pretrained GBDT model you also need Python packages from requirements.txt

How do I run it?

1. If you want to perform sanity check with euclidean distance you can download SIFT dataset and convert it to the appropriate format:

• cd data
• wget ftp://ftp.irisa.fr/local/texmex/corpus/sift.tar.gz
• tar -xvzf sift.tar.gz
• cd ..
• ./prepare --basesize 1000000 --base data/sift/sift_base.fvecs --querysize 10000 --query data/sift/sift_query.fvecs --trainsize 1000 --testsize 1000 --topsize 100 --outfolder data/sift --suffix sift

2. If you want to evaluate RPG on Collections and Video datasets, then download them using data/download.py script. If you want to download only precomputed relevance function scores, run:

• cd data
• python3 download.py all score

3. If you you want to download our precomputed GBDT models and the set of input features, run (note that it requires about 1 Tb of disk space and that you need only precomputed scores to reproduce our results):

• python3 download.py all model
• You can run these models using the demo-script data/compute_scores.py

4. After the data was prepared you need to build a graph and run the search algorithm:

• To perform experiments on SIFT dataset run:
  • ./RPG --mode base --baseSize 1000000 --trainQueries 1000 --base data/sift/train_sift.bin --outputGraph data/sift/graph.out --relevanceVector 100 --efConstruction 1000 --M 8
  • ./RPG --mode query --baseSize 1000000 --querySize 1000 --query data/sift/test_sift.bin --inputGraph data/sift/graph.out --efSearch 300 --topK 5 --output data/sift/result.txt --gtQueries 1000 --gtTop 100 --groundtruth data/sift/groundtruth_sift.bin
• To perform experiments on Collections dataset run:
  • ./RPG --mode base --baseSize 1000000 --trainQueries 1000 --base data/collections/data/model_scores/scores_train.bin --outputGraph data/collections/graph.out --relevanceVector 1000 --efConstruction 1000 --M 8
  • ./RPG --mode query --baseSize 1000000 --querySize 1000 --query data/collections/data/model_scores/scores_test.bin --inputGraph data/collections/graph.out --efSearch 300 --topK 5 --output data/collections/result.txt --gtQueries 1000 --gtTop 100 --groundtruth data/collections/data/model_scores/groundtruth.bin
• To perform experiments on Video dataset run:
  • ./RPG --mode base --baseSize 1000000 --trainQueries 1000 --base data/video/data/model_scores/scores_train.bin --outputGraph data/video/graph.out --relevanceVector 1000 --efConstruction 1000 --M 8
  • ./RPG --mode query --baseSize 1000000 --querySize 1000 --query data/video/data/model_scores/scores_test.bin --inputGraph data/video/graph.out --efSearch 300 --topK 5 --output data/video/result.txt --gtQueries 1000 --gtTop 100 --groundtruth data/video/data/model_scores/groundtruth.bin

5. You can vary efSearch to achieve higher recall values.