|
| 1 | +# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +import argparse |
| 16 | + |
| 17 | +from pipelines.document_stores import ElasticsearchDocumentStore |
| 18 | +from pipelines.nodes import ( |
| 19 | + BM25Retriever, |
| 20 | + DensePassageRetriever, |
| 21 | + ErnieRanker, |
| 22 | + JoinDocuments, |
| 23 | +) |
| 24 | +from pipelines.pipelines import Pipeline |
| 25 | +from pipelines.utils import ( |
| 26 | + convert_files_to_dicts, |
| 27 | + fetch_archive_from_http, |
| 28 | + print_documents, |
| 29 | +) |
| 30 | + |
| 31 | +# yapf: disable |
| 32 | +parser = argparse.ArgumentParser() |
| 33 | +parser.add_argument('--device', choices=['cpu', 'gpu'], default="gpu", help="Select which device to run dense_qa system, defaults to gpu.") |
| 34 | +parser.add_argument("--index_name", default='dureader_nano_query_encoder', type=str, help="The ann index name of ANN.") |
| 35 | +parser.add_argument("--search_engine", choices=['elastic'], default="elastic", help="The type of ANN search engine.") |
| 36 | +parser.add_argument("--max_seq_len_query", default=64, type=int, help="The maximum total length of query after tokenization.") |
| 37 | +parser.add_argument("--max_seq_len_passage", default=384, type=int, help="The maximum total length of passage after tokenization.") |
| 38 | +parser.add_argument("--retriever_batch_size", default=16, type=int, help="The batch size of retriever to extract passage embedding for building ANN index.") |
| 39 | +parser.add_argument("--query_embedding_model", default="rocketqa-zh-nano-query-encoder", type=str, help="The query_embedding_model path") |
| 40 | +parser.add_argument("--passage_embedding_model", default="rocketqa-zh-nano-para-encoder", type=str, help="The passage_embedding_model path") |
| 41 | +parser.add_argument("--params_path", default="", type=str, help="The checkpoint path") |
| 42 | +parser.add_argument("--embedding_dim", default=312, type=int, help="The embedding_dim of index") |
| 43 | +parser.add_argument('--host', type=str, default="localhost", help='host ip of ANN search engine') |
| 44 | +parser.add_argument('--port', type=str, default="9200", help='port of ANN search engine') |
| 45 | +parser.add_argument("--bm_topk", default=10, type=int, help="The number of candidates for BM25Retriever to retrieve.") |
| 46 | +parser.add_argument("--dense_topk", default=10, type=int, help="The number of candidates for DensePassageRetriever to retrieve.") |
| 47 | +parser.add_argument("--rank_topk", default=10, type=int, help="The number of candidates ranker to filter.") |
| 48 | + |
| 49 | +args = parser.parse_args() |
| 50 | +# yapf: enable |
| 51 | + |
| 52 | + |
| 53 | +def get_retrievers(use_gpu): |
| 54 | + |
| 55 | + doc_dir = "data/dureader_dev" |
| 56 | + dureader_data = "https://paddlenlp.bj.bcebos.com/applications/dureader_dev.zip" |
| 57 | + |
| 58 | + fetch_archive_from_http(url=dureader_data, output_dir=doc_dir) |
| 59 | + dicts = convert_files_to_dicts(dir_path=doc_dir, split_paragraphs=True, encoding="utf-8") |
| 60 | + |
| 61 | + document_store_with_docs = ElasticsearchDocumentStore( |
| 62 | + host=args.host, |
| 63 | + port=args.port, |
| 64 | + username="", |
| 65 | + password="", |
| 66 | + embedding_dim=312, |
| 67 | + search_fields=["content", "meta"], |
| 68 | + index=args.index_name, |
| 69 | + ) |
| 70 | + document_store_with_docs.write_documents(dicts) |
| 71 | + |
| 72 | + dpr_retriever = DensePassageRetriever( |
| 73 | + document_store=document_store_with_docs, |
| 74 | + query_embedding_model=args.query_embedding_model, |
| 75 | + passage_embedding_model=args.passage_embedding_model, |
| 76 | + params_path=args.params_path, |
| 77 | + output_emb_size=args.embedding_dim, |
| 78 | + max_seq_len_query=args.max_seq_len_query, |
| 79 | + max_seq_len_passage=args.max_seq_len_passage, |
| 80 | + batch_size=args.retriever_batch_size, |
| 81 | + use_gpu=use_gpu, |
| 82 | + embed_title=False, |
| 83 | + ) |
| 84 | + # update Embedding |
| 85 | + document_store_with_docs.update_embeddings(dpr_retriever) |
| 86 | + |
| 87 | + bm_retriever = BM25Retriever(document_store=document_store_with_docs) |
| 88 | + |
| 89 | + return dpr_retriever, bm_retriever |
| 90 | + |
| 91 | + |
| 92 | +def semantic_search_tutorial(): |
| 93 | + |
| 94 | + use_gpu = True if args.device == "gpu" else False |
| 95 | + |
| 96 | + dpr_retriever, bm_retriever = get_retrievers(use_gpu) |
| 97 | + |
| 98 | + # Ranker |
| 99 | + ranker = ErnieRanker(model_name_or_path="rocketqa-nano-cross-encoder", use_gpu=use_gpu) |
| 100 | + |
| 101 | + # Pipeline |
| 102 | + pipeline = Pipeline() |
| 103 | + pipeline.add_node(component=bm_retriever, name="BMRetriever", inputs=["Query"]) |
| 104 | + pipeline.add_node(component=dpr_retriever, name="DenseRetriever", inputs=["Query"]) |
| 105 | + pipeline.add_node( |
| 106 | + component=JoinDocuments(join_mode="concatenate"), name="JoinResults", inputs=["BMRetriever", "DenseRetriever"] |
| 107 | + ) |
| 108 | + pipeline.add_node(component=ranker, name="Ranker", inputs=["JoinResults"]) |
| 109 | + # Keywords recall results |
| 110 | + prediction = pipeline.run( |
| 111 | + query="广播权", params={"BMRetriever": {"top_k": 10}, "DenseRetriever": {"top_k": 10}, "Ranker": {"top_k": 3}} |
| 112 | + ) |
| 113 | + print_documents(prediction) |
| 114 | + # Dense vector recall results |
| 115 | + prediction = pipeline.run( |
| 116 | + query="期货交易手续费指的是什么?", |
| 117 | + params={"BMRetriever": {"top_k": 10}, "DenseRetriever": {"top_k": 10}, "Ranker": {"top_k": 3}}, |
| 118 | + ) |
| 119 | + pipeline.draw("multi_recall.png") |
| 120 | + print_documents(prediction) |
| 121 | + |
| 122 | + |
| 123 | +if __name__ == "__main__": |
| 124 | + semantic_search_tutorial() |
0 commit comments