DPR for OpenCSR
Please cite the original paper
@inproceedings{Karpukhin2020DensePR,
title={Dense Passage Retrieval for Open-Domain Question Answering},
author={V. Karpukhin and Barlas Oğuz and Sewon Min and Patrick Lewis and Ledell Yu Wu and Sergey Edunov and Danqi Chen and Wen-tau Yih},
booktitle={EMNLP},
year={2020}
}
Link to the code for the experiment: OpenCSR/baseline_methods/DPR/
Installation
conda create --name dpr python=3.7
conda activate dpr
# use our copy of the DPR code in `dpr-master`,
# or you can download the latest version (unverified) on Github
cd dpr-master
pip install -e dpr-master
pip install torch==1.4.0 torchvision==0.5.0 -f https://download.pytorch.org/whl/cu100/torch_stable.html
conda install faiss-gpu cudatoolkit=10.0 -c pytorch -n dpr
pip install transformers==3.0.2
pip install spacy
Preprocessing
Step 1. Transform the data format of the corpus.
You can skip this step. We have included the processed tsv file at gkb_best.prepro.tsv under drfact_data/knowledge_corpus/.
python -m language.labs.drfact.preprocessing.corpus_prepro --DATA_ROOT drfact_data/knowledge_corpus --CORPUS_PATH GenericsKB-Best.tsv --OUTPUT_JSON_PATH gkb_best.prepro.jsonl
python baseline_methods/DPR/convert_gkb_tsv.py ${CORPUS_PATH}/gkb_best.prepro.jsonl
# This will generate `${CORPUS_PATH}/gkb_best.prepro.tsv`
Step 2. Convert the OpenCSR datasets format.
declare -a datasets=("ARC" "OBQA" "QASC")
declare -a splits=("train" "dev")
for DATA_NAME in "${datasets[@]}"
do
DATA_FOLDER=drfact_data/datasets/${DATA_NAME}
python baseline_methods/DPR/convert_qas_csv.py $DATA_FOLDER/linked_train.jsonl no &
python baseline_methods/DPR/convert_qas_csv.py $DATA_FOLDER/linked_dev.jsonl no &
done
Step 3. Generate data from BM25 for training DPR
Note that you need to do the inference of BM25 first.
declare -a datasets=("ARC" "OBQA" "QASC")
declare -a splits=("train" "dev")
for DATA_NAME in "${datasets[@]}"
do
DATA_FOLDER=drfact_data/datasets/${DATA_NAME}
for SPLIT in "${splits[@]}"
do
python baseline_methods/DPR/genenrate_dpr_data.py \
--dataset_name ${DATA_NAME} \
--linked_qas_file ${DATA_FOLDER}/linked_${SPLIT}.jsonl \
--drfact_format_gkb_file ${CORPUS_PATH}/gkb_best.drfact_format.jsonl \
--ret_result_file ${DATA_FOLDER}/linked_${SPLIT}.BM25.jsonl \
--output_file baseline_methods/DPR/dpr_train_data/${DATA_NAME}_${SPLIT}.dpr_format.jsonl &
done
done
# After the reformatting, we get the json format instead of the jsonl format
declare -a datasets=("ARC" "OBQA" "QASC")
declare -a splits=("train" "dev")
RET=BM25
for DATA_NAME in "${datasets[@]}"
do
DATA_FOLDER=drfact_data/datasets/${DATA_NAME}
for SPLIT in "${splits[@]}"
do
python baseline_methods/DPR/jsonl2json.py \
baseline_methods/DPR/dpr_train_data/${DATA_NAME}_${SPLIT}.dpr_format.jsonl &
done
done
Training DPR
Step 4. Train DPR index for each dataset.
mkdir ~/dpr_data/ # Can be any folder but need to be consistent.
DPR_DATA=baseline_methods/DPR/dpr_train_data/
declare -a datasets=("ARC" "OBQA" "QASC")
for DATA_NAME in "${datasets[@]}"
do
echo "${DATA_NAME}"
gpu=0,1,2,3,4,5,6,7
master_port=29500
nproc_per_node=8 # Note this should be correct.
echo "CUDA_VISIBLE_DEVICES=${gpu}"
rm -r ~/dpr_data/${DATA_NAME}_dpr_model/
mkdir ~/dpr_data/${DATA_NAME}_dpr_model/
CUDA_VISIBLE_DEVICES=${gpu} python -m torch.distributed.launch \
--master_port ${master_port} \
--nproc_per_node=${nproc_per_node} \
baseline_methods/DPR/dpr-master/train_dense_encoder.py \
--max_grad_norm 2.0 \
--encoder_model_type hf_bert \
--pretrained_model_cfg bert-base-uncased \
--seed 12345 \
--sequence_length 64 \
--warmup_steps 1237 \
--batch_size 16 \
--do_lower_case \
--train_file ${DPR_DATA}/${DATA_NAME}_train.dpr_format.json \
--dev_file ${DPR_DATA}/${DATA_NAME}_dev.dpr_format.json \
--output_dir ~/dpr_data/${DATA_NAME}_dpr_model/ \
--learning_rate 3e-05 \
--num_train_epochs 10 \
--dev_batch_size 16 \
--val_av_rank_start_epoch 10 --fp16
done
Inference
Step 5. Encode facts with the trained DPR model for each dataset.
gpu=0
declare -a datasets=("ARC" "OBQA" "QASC")
for DATA_NAME in "${datasets[@]}"
do
DPR_FILE=$(ls ~/dpr_data/${DATA_NAME}_dpr_model/ -Art | tail -n 1) # use the last checkpoint.
((gpup=gpu+1))
echo "CUDA_VISIBLE_DEVICES=${gpu},${gpup}: ${DPR_FILE}"
CUDA_VISIBLE_DEVICES=${gpu},${gpup} python baseline_methods/DPR/dpr-master/generate_dense_embeddings.py \
--model_file ~/dpr_data/${DATA_NAME}_dpr_model/${DPR_FILE} \
--ctx_file ~/drfact/drfact_data/knowledge_corpus/gkb_best.prepro.tsv \
--batch_size 5000 --shard_id 0 --num_shards 1 \
--out_file ~/drfact/drfact_data/${DATA_NAME}_dpr_index &
((gpu=gpu+2))
done
Step 6. DPR Inference for each dataset
declare -a datasets=("ARC" "OBQA" "QASC")
declare -a splits=("train" "dev" "test")
gpu=0
for DATA_NAME in "${datasets[@]}"
do
DATA_FOLDER=drfact_data/datasets/${DATA_NAME}
for SPLIT in "${splits[@]}"
do
echo "${gpu}: ${DATA_NAME}-${SPLIT}"
DPR_FILE=$(ls ~/dpr_data/${DATA_NAME}_dpr_model/ -Art | tail -n 1)
CUDA_VISIBLE_DEVICES=${gpu} python baseline_methods/DPR/dense_retriever.py \
--model_file ~/dpr_data/${DATA_NAME}_dpr_model/${DPR_FILE} \
--ctx_file ~/drfact/drfact_data/knowledge_corpus/gkb_best.prepro.tsv \
--qa_file $DATA_FOLDER/linked_${SPLIT}.csv \
--encoded_ctx_file ~/drfact/drfact_data/${DATA_NAME}_dpr_index_0.pkl \
--out_file ~/dpr_data/${DATA_NAME}-${SPLIT}_${DATA_NAME}-DPR_result_1000.pkl \
--batch_size 512 --n-docs 1000 &
((gpu=gpu+1))
done
done
Step 7. DPR Retrieval Formatting for each dataset
declare -a datasets=("ARC" "OBQA" "QASC")
declare -a splits=("train" "dev" "test")
for DATA_NAME in "${datasets[@]}"
do
DATA_FOLDER=drfact_data/datasets/${DATA_NAME}
for SPLIT in "${splits[@]}"
do
python baseline_methods/DPR/dpr_result_formatter.py \
--linked_qa_file ${DATA_FOLDER}/linked_${SPLIT}.jsonl \
--output_prefix ${DATA_NAME}_DPR \
--drfact_format_gkb_file ${CORPUS_PATH}/gkb_best.drfact_format.jsonl \
--dpr_result_file ~/dpr_data/${DATA_NAME}-${SPLIT}_${DATA_NAME}-DPR_result_1000.pkl &
done
done