def init_query_encoder(encoder, topics_name, encoded_queries, device):
encoded_queries_map = {
'msmarco-passage-dev-subset': 'tct_colbert-msmarco-passage-dev-subset',
'dpr-nq-dev': 'dpr_multi-nq-dev',
'dpr-nq-test': 'dpr_multi-nq-test',
'dpr-trivia-dev': 'dpr_multi-trivia-dev',
'dpr-trivia-test': 'dpr_multi-trivia-test',
'dpr-wq-test': 'dpr_multi-wq-test',
'dpr-squad-test': 'dpr_multi-squad-test',
'dpr-curated-test': 'dpr_multi-curated-test'
}
if encoder:
if 'dpr' in encoder:
return DprQueryEncoder(encoder_dir=encoder, device=device)
elif 'tct_colbert' in encoder:
return TctColBertQueryEncoder(encoder_dir=encoder, device=device)
elif 'ance' in encoder:
return AnceQueryEncoder(encoder_dir=encoder, device=device)
elif 'sentence' in encoder:
return AutoQueryEncoder(encoder_dir=encoder,
device=device,
pooling='mean',
l2_norm=True)
else:
return AutoQueryEncoder(encoder_dir=encoder, device=device)
if encoded_queries:
if os.path.exists(encoded_queries):
return QueryEncoder(encoded_queries)
return QueryEncoder.load_encoded_queries(encoded_queries)
if topics_name in encoded_queries_map:
return QueryEncoder.load_encoded_queries(
encoded_queries_map[topics_name])
return None
def init_query_encoder(encoder, tokenizer_name, topics_name, encoded_queries,
device, prefix):
encoded_queries_map = {
'msmarco-passage-dev-subset': 'tct_colbert-msmarco-passage-dev-subset',
'dpr-nq-dev': 'dpr_multi-nq-dev',
'dpr-nq-test': 'dpr_multi-nq-test',
'dpr-trivia-dev': 'dpr_multi-trivia-dev',
'dpr-trivia-test': 'dpr_multi-trivia-test',
'dpr-wq-test': 'dpr_multi-wq-test',
'dpr-squad-test': 'dpr_multi-squad-test',
'dpr-curated-test': 'dpr_multi-curated-test'
}
if encoder:
if 'dkrr' in encoder:
return DkrrDprQueryEncoder(encoder_dir=encoder,
device=device,
prefix=prefix)
elif 'dpr' in encoder:
return DprQueryEncoder(encoder_dir=encoder,
tokenizer_name=tokenizer_name,
device=device)
elif 'bpr' in encoder:
return BprQueryEncoder(encoder_dir=encoder,
tokenizer_name=tokenizer_name,
device=device)
elif 'tct_colbert' in encoder:
return TctColBertQueryEncoder(encoder_dir=encoder,
tokenizer_name=tokenizer_name,
device=device)
elif 'ance' in encoder:
return AnceQueryEncoder(encoder_dir=encoder,
tokenizer_name=tokenizer_name,
device=device)
elif 'sentence' in encoder:
return AutoQueryEncoder(encoder_dir=encoder,
tokenizer_name=tokenizer_name,
device=device,
pooling='mean',
l2_norm=True)
else:
return AutoQueryEncoder(encoder_dir=encoder,
tokenizer_name=tokenizer_name,
device=device)
if encoded_queries:
if os.path.exists(encoded_queries):
if 'bpr' in encoded_queries:
return BprQueryEncoder(encoded_query_dir=encoded_queries)
else:
return QueryEncoder(encoded_queries)
return QueryEncoder.load_encoded_queries(encoded_queries)
if topics_name in encoded_queries_map:
return QueryEncoder.load_encoded_queries(
encoded_queries_map[topics_name])
raise ValueError(f'No encoded queries for topic {topics_name}')
def init_query_encoder(encoder, topics_name, encoded_queries, device):
encoded_queries_map = {
'msmarco-passage-dev-subset': 'msmarco-passage-dev-subset-tct_colbert',
'dpr-nq-dev': 'dpr-nq-dev-multi',
'dpr-nq-test': 'dpr-nq-test-multi',
'dpr-trivia-dev': 'dpr-trivia-dev-multi',
'dpr-trivia-test': 'dpr-trivia-test-multi',
'dpr-wq-test': 'dpr-wq-test-multi',
'dpr-squad-test': 'dpr-squad-test-multi',
'dpr-curated-test': 'dpr-curated-test-multi'
}
if encoder:
if 'dpr' in encoder:
return DPRQueryEncoder(encoder_dir=encoder, device=device)
elif 'tct_colbert' in encoder:
return TCTColBERTQueryEncoder(encoder_dir=encoder, device=device)
elif 'ance' in encoder:
return AnceQueryEncoder(encoder_dir=encoder, device=device)
elif 'sentence' in encoder:
return SBERTQueryEncoder(encoder_dir=encoder, device=device)
else:
return AutoQueryEncoder(encoder_dir=encoder, device=device)
if encoded_queries:
if os.path.exists(encoded_queries):
return QueryEncoder(encoded_queries)
return QueryEncoder.load_encoded_queries(encoded_queries)
if topics_name in encoded_queries_map:
return QueryEncoder.load_encoded_queries(
encoded_queries_map[topics_name])
return None
parser.add_argument('--output', type=str, help='dir to store query embeddings', required=True)
parser.add_argument('--device', type=str,
help='device cpu or cuda [cuda:0, cuda:1...]', default='cpu', required=False)
args = parser.parse_args()
device = args.device
topics = get_topics(args.topics)
if not os.path.exists(args.output):
os.mkdir(args.output)
if 'dpr' in args.encoder:
encoder = DprQueryEncoder(encoder_dir=args.encoder, device=device)
elif 'tct_colbert' in args.encoder:
encoder = TctColBertQueryEncoder(encoder_dir=args.encoder, device=device)
elif 'ance' in args.encoder:
encoder = AnceQueryEncoder(encoder_dir=args.encoder, device=device)
elif 'sentence' in args.encoder:
encoder = AutoQueryEncoder(encoder_dir=args.encoder, device=device, pooling='mean', l2_norm=True)
else:
encoder = AutoQueryEncoder(encoder_dir=args.encoder, device=device)
embeddings = {'id': [], 'text': [], 'embedding': []}
for key in tqdm(topics):
qid = str(key)
text = topics[key]['title']
embeddings['id'].append(qid)
embeddings['text'].append(text)
embeddings['embedding'].append(encoder.encode(text.strip()))
embeddings = pd.DataFrame(embeddings)
embeddings.to_pickle(os.path.join(args.output, 'embedding.pkl'))