annbench is a simple benchmark for approximate nearest neighbor search algorithms in Python. This repository design is strongly influenced by a great project, ann-benchmarks, that provides comprehensive and thorough benchmarks for various algorithms. In contrast, we aim to deliver more lightweight and straightforward scripts with the following features.

• Support Euclidean distance only
• Support Recall@1 only
• Support libraries installable via pip/conda only
• Search with a single thread
• Sweep by a single query parameter

• AWS Deep Learning AMI, c5.4xlarge (Aug 28, 2021)

git clone https://github.com/matsui528/annbench.git
cd annbench
pip install -r requirements.txt
# conda install faiss-cpu -y -c pytorch  # If you'd like to try faiss, run this on anaconda
# conda install faiss-gpu -y -c pytorch  # or, if you have GPUs, install faiss-gpu

python download.py dataset=siftsmall  # Downloaded on ./dataset

python run.py dataset=siftsmall algo=annoy  # Indices are on ./interim. Results are on ./output

python plot.py   # Plots are on ./result_img

# Downloading deep1m takes some hours
python download.py --multirun dataset=siftsmall,sift1m,deep1m

# Will take some hours
python run.py --multirun dataset=siftsmall,sift1m,deep1m algo=linear,annoy,ivfpq,hnsw,ivfpq4bit,scann,pq,pq4bit,hnsw_faiss,nsg
# Or, if you have GPUs, 
# python run.py --multirun dataset=siftsmall,sift1m,deep1m algo=linear,annoy,ivfpq,hnsw,linear_gpu,ivfpq_gpu,ivfpq4bit,scann,pq,pq4bit,hnsw_faiss,nsg

python plot.py

• All config files are on ./conf. You can edit the config files to change parameters.

• Download a target dataset by python download.py dataset=DATASET. Several datasets can be downloaded at once by python download.py --multirun dataset=DATASET1,DATASET2,DATASET3. See hydra for more detailed APIs for multirun.
• The data will be placed on ./dataset.

• Evaluate a target algorithm (ALGO) with a target dataset (DATASET) by python run.py dataset=DATASET algo=ALGO. You can run multiple algorithms on multiple datasets by python run.py --multirun dataset=DATASET1,DATASET2 algo=ALGO1,ALGO2.
• Indices (data structures for search) are stored on ./interim. They are reused for each run with different query parameters.
• The search result will be on ./output.
• By default, we run each algorithm num_trial=10 times and return the average runtime. You can change this by: python run.py num_trial=5

• You can visualize the search result by python plot.py. This script checks ./output and generate figures for each dataset on ./result_img.
• In order not to print query parameters, you can set the flag false: python plot.py with_query_param=false.
• To change the size of the image: python plot.py width=15 height=10

• When running run.py or plot.py, the output files will be on ./log as well. For example with python run.py algo=annoy dataset=siftsmall, the result file will be saved on (1) ./output/siftsmall/annoy/result.yaml and (2) ./log/2020-03-11/22-30-59/0/result.yaml.

• linear scan (faiss)
• pq (faiss)
• 4-bit pq(faiss)
• ivfpq (faiss)
• ivfpq with 4-bit quantizer (faiss)
• hnsw (faiss)
• nsg (faiss)
• annoy
• hnsw (hnswlib)
• scann
• pynndescent

Note that hnsw (hnswlib) is an original implementation by the original authors, and hnsw (faiss) is a re-implemented version by the faiss team.

• Write a wrapper class on ./annbench/algo. The class must inherit BaseANN class. See annoy.py for examples.
• Update ./annbench/algo/proxy.py
• Add the name of the library on requirements.txt.
• Add a config file on ./conf/algo.
• Make sure the algorithm runs on a single thread

• Write a wrapper class that inherits BaseDataset on ./annbench/dataset. An simple example is siftsmall.py.
• Update ./annbench/dataset/proxy.py.
• Add a config file on ./conf/dataset.

• We followed the standard evaluation criteria in the field of computer vision, Recall@1. See the evaluation function for more details. These functions are from the benchmark scripts of the faiss library.

• We define a simple guideline to set parameters. An algorithm has to have several index parameters and a single query parameter. For one set of index parameters, one index (data structure) is built. For this index, we run the search by sweeping the query parameter.
• For example with ivfpq, let us consider the following index parameters:

  param_index={"M": 8, "nlist": 100}

  With these parameters, one index (let us denote ivfpq(M=8, nlist=100)) is created. This index is stored in the disk as M8_nlist100.bin, where the way of naming is defined in the function stringify_index_param. Here, a query parameter is defined as:

  param_query={"nprobe": [1, 2, 4, 8, 16]}

  In the search step, the index is read from the disk onto the memory first. Then we run the search five times, with for nprobe in [1, 2, 4, 8, 16]. This creates five results (five pairs of (recall, runtime)). By connecting these results, one polyline is drawn on the final plot.
• Note that you must sort the values of the above query parameter. If you forget to sort (e.g., [1, 4, 2, 8, 16]), the final graph would become weird.

• Index/query parameters for each algorithm is defined in ./conf/algo/. These parameters are used for all datasets by default. If you'd like to specialize parameters for a specific dataset, you can define the specialized version in ./conf/specialized_param/.
• For example, the default parameters for ivfpq is defined here, where nlist=100. You can set nlist=1000 for the sift1m dataset by adding a config file here

• In addition to editing the config files, you can override values from the commandline thanks to hydra.
• Examples:
  • Writing index structures on SOMEWHEE/LARGE_HDD/:

    python run.py interim=SOMEWHERE/LARGE_HDD/interim

  • Run the ivfpq algorithm with different query parameters:

    python run.py algo=ivfpq dataset=siftsmall param_query.nprobe=[1,5,25]

• ANN Benchmarks

Feel free to open a pull request.

• @foreverYoungGitHub (hnsw_faiss and nsg)

• Yusuke Matsui