TACO

Official implementation of TACO paper. Our code is built on Openmatch.

Setup

pip install -r requirements.txt
pip install -e .

Install Faiss-gpu

For A100 CUDA-11, we need to install faiss-gpu via conda or specific pip wheels

• via conda

# check your cudatoolkit
conda search cudatoolkit
# install faiss-gpu 
conda install -c pytorch faiss-gpu cudatoolkit=11.3.1

• via pip

# download faiss-gpu pip cuda-11 wheel
wget https://github.com/kyamagu/faiss-wheels/releases/download/v1.7.3/faiss_gpu-1.7.3-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

# pip install
python -m pip install faiss_gpu-1.7.3-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

Supported Datasets and Setup DATA

• supported datasets
  • KILT Benchmark
  • MSMARCO
  • ZESHEL
  • BEIR
• setup data scripts (kilt for example)

./scripts/kilt/shells/setup_data.sh

For ZESHEL, you need to download the raw data and bm25 candidates from here , then run the setup_data.sh script

Supported features:

• standard dense retrieval
  • hard negative mining: unify ANCE with Hard-nce style, the only difference is optimizer state
• multi-task retrieval
  • PCG
  • CGD
  • GradNorm
  • TACO
  • Naive
  • split query encoder
  • query adapter

example Scripts

• T5-ANCE dense retrieval

# warmup with bm25 candidates first
./scripts/kilt/shells/warmup_dr.sh

# ANCE iterations
./scripts/kilt/shells/ance_dr.sh

• Multi_task dense retrieval

# warmup with bm25 candidates first
./scripts/kilt/shells/warmup_mt.sh 

# ANCE iterations
./scripts/kilt/shells/ance_mt.sh [your multi_task method (naive, pcg, gn, cgd
, taco]