Add Ernie Search base model into pipelines (#3906)

w5688414 · web-flow · commit 930a986f4a46 · 2022-11-28T10:02:59.000+08:00
* Add Ernie Search base version into pipelines

* Add more comments
diff --git a/paddlenlp/transformers/ernie/modeling.py b/paddlenlp/transformers/ernie/modeling.py
@@ -715,6 +715,33 @@ class ErniePretrainedModel(PretrainedModel):
             "vocab_size": 30522,
             "pad_token_id": 0,
         },
+        "ernie-search-base-dual-encoder-marco-en": {
+            "attention_probs_dropout_prob": 0.1,
+            "hidden_act": "gelu",
+            "hidden_dropout_prob": 0.1,
+            "hidden_size": 768,
+            "initializer_range": 0.02,
+            "max_position_embeddings": 512,
+            "num_attention_heads": 12,
+            "num_hidden_layers": 12,
+            "type_vocab_size": 4,
+            "vocab_size": 30522,
+            "pad_token_id": 0,
+        },
+        "ernie-search-large-cross-encoder-marco-en": {
+            "attention_probs_dropout_prob": 0.1,
+            "intermediate_size": 4096,
+            "hidden_act": "gelu",
+            "hidden_dropout_prob": 0.1,
+            "hidden_size": 1024,
+            "initializer_range": 0.02,
+            "max_position_embeddings": 512,
+            "num_attention_heads": 16,
+            "num_hidden_layers": 24,
+            "type_vocab_size": 4,
+            "vocab_size": 30522,
+            "pad_token_id": 0,
+        },
     }
     resource_files_names = {"model_state": "model_state.pdparams"}
     pretrained_resource_files_map = {
@@ -800,6 +827,10 @@ class ErniePretrainedModel(PretrainedModel):
             "https://paddlenlp.bj.bcebos.com/models/transformers/rocketqa/rocketqav2_en_marco_query_encoder.pdparams",
             "rocketqav2-en-marco-para-encoder":
             "https://paddlenlp.bj.bcebos.com/models/transformers/rocketqa/rocketqav2_en_marco_para_encoder.pdparams",
+            "ernie-search-base-dual-encoder-marco-en":
+            "https://paddlenlp.bj.bcebos.com/models/transformers/ernie_search/ernie_search_base_dual_encoder_marco_en.pdparams",
+            "ernie-search-large-cross-encoder-marco-en":
+            "https://paddlenlp.bj.bcebos.com/models/transformers/ernie_search/ernie_search_large_cross_encoder_marco_en.pdparams",
         }
     }
     base_model_prefix = "ernie"
diff --git a/paddlenlp/transformers/ernie/tokenizer.py b/paddlenlp/transformers/ernie/tokenizer.py
@@ -70,6 +70,8 @@
     "rocketqav2-en-marco-cross-encoder": 512,
     "rocketqav2-en-marco-query-encoder": 512,
     "rocketqav2-en-marco-para-encoder": 512,
+    "ernie-search-base-dual-encoder-marco-en": 512,
+    "ernie-search-large-cross-encoder-marco-en": 512,
 }
 
 
@@ -211,6 +213,10 @@ class ErnieTokenizer(PretrainedTokenizer):
             "https://bj.bcebos.com/paddlenlp/models/transformers/ernie_v2_base/vocab.txt",
             "rocketqav2-en-marco-para-encoder":
             "https://bj.bcebos.com/paddlenlp/models/transformers/ernie_v2_base/vocab.txt",
+            "ernie-search-base-dual-encoder-marco-en":
+            "https://bj.bcebos.com/paddlenlp/models/transformers/ernie_v2_base/vocab.txt",
+            "ernie-search-large-cross-encoder-marco-en":
+            "https://bj.bcebos.com/paddlenlp/models/transformers/ernie_v2_large/vocab.txt",
         }
     }
     pretrained_init_configuration = {
@@ -343,6 +349,12 @@ class ErnieTokenizer(PretrainedTokenizer):
         "rocketqav2-en-marco-para-encoder": {
             "do_lower_case": True
         },
+        "ernie-search-base-dual-encoder-marco-en": {
+            "do_lower_case": True
+        },
+        "ernie-search-large-cross-encoder-marco-en": {
+            "do_lower_case": True
+        },
     }
 
     max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
diff --git a/paddlenlp/transformers/semantic_search/modeling.py b/paddlenlp/transformers/semantic_search/modeling.py
@@ -23,12 +23,25 @@
 
 class ErnieEncoder(ErniePretrainedModel):
 
-    def __init__(self, ernie, dropout=None, num_classes=2):
+    def __init__(self,
+                 ernie,
+                 dropout=None,
+                 output_emb_size=None,
+                 num_classes=2):
         super(ErnieEncoder, self).__init__()
         self.ernie = ernie  # allow ernie to be config
         self.dropout = nn.Dropout(dropout if dropout is not None else 0.1)
         self.classifier = nn.Linear(self.ernie.config["hidden_size"],
                                     num_classes)
+        # Compatible to ERNIE-Search for adding extra linear layer
+        self.output_emb_size = output_emb_size
+        if output_emb_size is not None and output_emb_size > 0:
+            weight_attr = paddle.ParamAttr(
+                initializer=paddle.nn.initializer.TruncatedNormal(std=0.02))
+            self.emb_reduce_linear = paddle.nn.Linear(
+                self.ernie.config["hidden_size"],
+                output_emb_size,
+                weight_attr=weight_attr)
         self.apply(self.init_weights)
 
     def init_weights(self, layer):
@@ -79,21 +92,23 @@ def __init__(self,
                  query_model_name_or_path=None,
                  title_model_name_or_path=None,
                  share_parameters=False,
+                 output_emb_size=None,
                  dropout=None,
                  reinitialize=False,
                  use_cross_batch=False):
 
         super().__init__()
         self.query_ernie, self.title_ernie = None, None
         self.use_cross_batch = use_cross_batch
+        self.output_emb_size = output_emb_size
         if query_model_name_or_path is not None:
             self.query_ernie = ErnieEncoder.from_pretrained(
-                query_model_name_or_path)
+                query_model_name_or_path, output_emb_size=output_emb_size)
         if share_parameters:
             self.title_ernie = self.query_ernie
         elif title_model_name_or_path is not None:
             self.title_ernie = ErnieEncoder.from_pretrained(
-                title_model_name_or_path)
+                title_model_name_or_path, output_emb_size=output_emb_size)
         assert (self.query_ernie is not None) or (self.title_ernie is not None), \
             "At least one of query_ernie and title_ernie should not be None"
 
@@ -125,16 +140,27 @@ def get_pooled_embedding(self,
                              position_ids=None,
                              attention_mask=None,
                              is_query=True):
+        """Get the first feature of each sequence for classification"""
         assert (is_query and self.query_ernie is not None) or (not is_query and self.title_ernie), \
             "Please check whether your parameter for `is_query` are consistent with DualEncoder initialization."
         if is_query:
             sequence_output, _ = self.query_ernie(input_ids, token_type_ids,
                                                   position_ids, attention_mask)
+            if self.output_emb_size is not None and self.output_emb_size > 0:
+                cls_embedding = self.query_ernie.emb_reduce_linear(
+                    sequence_output[:, 0])
+            else:
+                cls_embedding = sequence_output[:, 0]
 
         else:
             sequence_output, _ = self.title_ernie(input_ids, token_type_ids,
                                                   position_ids, attention_mask)
-        return sequence_output[:, 0]
+            if self.output_emb_size is not None and self.output_emb_size > 0:
+                cls_embedding = self.title_ernie.emb_reduce_linear(
+                    sequence_output[:, 0])
+            else:
+                cls_embedding = sequence_output[:, 0]
+        return cls_embedding
 
     def cosine_sim(self,
                    query_input_ids,
@@ -272,6 +298,7 @@ def matching(self,
                  position_ids=None,
                  attention_mask=None,
                  return_prob_distributation=False):
+        """Use the pooled_output as the feature for pointwise prediction, eg. RocketQAv1"""
         _, pooled_output = self.ernie(input_ids,
                                       token_type_ids=token_type_ids,
                                       position_ids=position_ids,
@@ -288,6 +315,7 @@ def matching_v2(self,
                     token_type_ids=None,
                     position_ids=None,
                     attention_mask=None):
+        """Use the cls token embedding as the feature for listwise prediction, eg. RocketQAv2"""
         sequence_output, _ = self.ernie(input_ids,
                                         token_type_ids=token_type_ids,
                                         position_ids=position_ids,
@@ -296,6 +324,21 @@ def matching_v2(self,
         probs = self.ernie.classifier(pooled_output)
         return probs
 
+    def matching_v3(self,
+                    input_ids,
+                    token_type_ids=None,
+                    position_ids=None,
+                    attention_mask=None):
+        """Use the pooled_output as the feature for listwise prediction, eg. ERNIE-Search"""
+        sequence_output, pooled_output = self.ernie(
+            input_ids,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask)
+        pooled_output = self.ernie.dropout(pooled_output)
+        probs = self.ernie.classifier(pooled_output)
+        return probs
+
     def forward(self,
                 input_ids,
                 token_type_ids=None,
diff --git a/pipelines/API.md b/pipelines/API.md
@@ -14,6 +14,7 @@
 | rocketqa-zh-micro-query-encoder    | Chinese     | 4-layer, 384-hidden, 12-heads, 23M parameters. Trained on DuReader retrieval text.     |
 | rocketqa-zh-nano-query-encoder     | Chinese     | 4-layer, 312-hidden, 12-heads, 18M parameters. Trained on DuReader retrieval text.     |
 | rocketqav2-en-marco-query-encoder    | English     | 12-layer, 768-hidden, 12-heads, 118M parameters. Trained on MSMARCO.     |
+| ernie-search-base-dual-encoder-marco-en    | English     | 12-layer, 768-hidden, 12-heads, 118M parameters. Trained on MSMARCO.     |
 
 ## ErnieRanker
 
@@ -27,6 +28,7 @@
 | rocketqa-micro-cross-encoder     | Chinese     | 4-layer, 384-hidden, 12-heads, 23M parameters. Trained on DuReader retrieval text.     |
 | rocketqa-nano-cross-encoder    | Chinese     | 4-layer, 312-hidden, 12-heads, 18M parameters. Trained on DuReader retrieval text.    |
 | rocketqav2-en-marco-cross-encoder    | English     | 12-layer, 768-hidden, 12-heads, 118M parameters. Trained on Trained on MSMARCO.    |
+| ernie-search-large-cross-encoder-marco-en    | English     | 24-layer, 768-hidden, 12-heads, 118M parameters. Trained on Trained on MSMARCO.    |
 
 ## ErnieReader
 
diff --git a/pipelines/pipelines/nodes/retriever/dense.py b/pipelines/pipelines/nodes/retriever/dense.py
@@ -49,7 +49,9 @@ def __init__(
             Path, str] = "rocketqa-zh-dureader-para-encoder",
         params_path: Optional[str] = "",
         model_version: Optional[str] = None,
-        output_emb_size=256,
+        output_emb_size: Optional[int] = None,
+        reinitialize: bool = False,
+        share_parameters: bool = False,
         max_seq_len_query: int = 64,
         max_seq_len_passage: int = 256,
         top_k: int = 10,
@@ -98,7 +100,7 @@ def __init__(
         :param progress_bar: Whether to show a tqdm progress bar or not.
                              Can be helpful to disable in production deployments to keep the logs clean.
         """
-        # save init parameters to enable export of component config as YAML
+        # Save init parameters to enable export of component config as YAML
         self.set_config(
             document_store=document_store,
             query_embedding_model=query_embedding_model,
@@ -110,6 +112,9 @@ def __init__(
             use_gpu=use_gpu,
             batch_size=batch_size,
             embed_title=embed_title,
+            reinitialize=reinitialize,
+            share_parameters=share_parameters,
+            output_emb_size=output_emb_size,
             similarity_function=similarity_function,
             progress_bar=progress_bar,
         )
@@ -150,8 +155,12 @@ def __init__(
             self.passage_tokenizer = AutoTokenizer.from_pretrained(
                 query_embedding_model)
         else:
-            self.ernie_dual_encoder = ErnieDualEncoder(query_embedding_model,
-                                                       passage_embedding_model)
+            self.ernie_dual_encoder = ErnieDualEncoder(
+                query_embedding_model,
+                passage_embedding_model,
+                output_emb_size=output_emb_size,
+                reinitialize=reinitialize,
+                share_parameters=share_parameters)
             self.query_tokenizer = AutoTokenizer.from_pretrained(
                 query_embedding_model)
             self.passage_tokenizer = AutoTokenizer.from_pretrained(
diff --git a/pipelines/rest_api/config.py b/pipelines/rest_api/config.py
@@ -19,6 +19,7 @@
     "PIPELINE_YAML_PATH",
     str((Path(__file__).parent / "pipeline" / "pipelines.yaml").absolute()))
 QUERY_PIPELINE_NAME = os.getenv("QUERY_PIPELINE_NAME", "query")
+QUERY_QA_PAIRS_NAME = os.getenv('QUERY_QA_PAIRS_NAME', 'query_qa_pairs')
 INDEXING_PIPELINE_NAME = os.getenv("INDEXING_PIPELINE_NAME", "indexing")
 INDEXING_QA_GENERATING_PIPELINE_NAME = os.getenv(
     "INDEXING_QA_GENERATING_PIPELINE_NAME", "indexing_qa_generating")
diff --git a/pipelines/rest_api/controller/search.py b/pipelines/rest_api/controller/search.py
@@ -25,7 +25,7 @@
 
 import pipelines
 from pipelines.pipelines.base import Pipeline
-from rest_api.config import PIPELINE_YAML_PATH, QUERY_PIPELINE_NAME
+from rest_api.config import PIPELINE_YAML_PATH, QUERY_PIPELINE_NAME, QUERY_QA_PAIRS_NAME
 from rest_api.config import LOG_LEVEL, CONCURRENT_REQUEST_PER_WORKER
 from rest_api.schema import QueryRequest, QueryResponse, DocumentRequest, DocumentResponse, QueryImageResponse, QueryQAPairResponse, QueryQAPairRequest
 from rest_api.controller.utils import RequestLimiter
@@ -42,8 +42,11 @@
 PIPELINE = Pipeline.load_from_yaml(Path(PIPELINE_YAML_PATH),
                                    pipeline_name=QUERY_PIPELINE_NAME)
 
-QA_PAIR_PIPELINE = Pipeline.load_from_yaml(Path(PIPELINE_YAML_PATH),
-                                           pipeline_name="query_qa_pairs")
+try:
+    QA_PAIR_PIPELINE = Pipeline.load_from_yaml(
+        Path(PIPELINE_YAML_PATH), pipeline_name=QUERY_QA_PAIRS_NAME)
+except Exception as e:
+    logger.warning(f"Request pipeline ('{QUERY_QA_PAIRS_NAME}: is null'). ")
 DOCUMENT_STORE = PIPELINE.get_document_store()
 logging.info(f"Loaded pipeline nodes: {PIPELINE.graph.nodes.keys()}")
 
diff --git a/pipelines/rest_api/pipeline/semantic_ernie_search_en.yaml b/pipelines/rest_api/pipeline/semantic_ernie_search_en.yaml
@@ -0,0 +1,69 @@
+version: '1.1.0'
+
+components:    # define all the building-blocks for Pipeline
+  - name: DocumentStore
+    type: ElasticsearchDocumentStore  # consider using Milvus2DocumentStore or WeaviateDocumentStore for scaling to large number of documents
+    params:
+      host: localhost
+      port: 9200
+      index: msmarco
+      embedding_dim: 768
+  - name: Retriever
+    type: DensePassageRetriever
+    params:
+      document_store: DocumentStore    # params can reference other components defined in the YAML
+      top_k: 10
+      query_embedding_model: ernie-search-base-dual-encoder-marco-en # an example of using ernie search models
+      share_parameters: True
+      output_emb_size: 768
+      embed_title: False
+  - name: Ranker       # custom-name for the component; helpful for visualization & debugging
+    type: ErnieRanker    # pipelines Class name for the component
+    params:
+      model_name_or_path: rocketqav2-en-marco-cross-encoder
+      top_k: 3
+      use_en: True,
+      reinitialize: True
+  - name: TextFileConverter
+    type: TextConverter
+  - name: ImageFileConverter
+    type: ImageToTextConverter
+  - name: PDFFileConverter
+    type: PDFToTextConverter
+  - name: DocxFileConverter
+    type: DocxToTextConverter
+  - name: Preprocessor
+    type: PreProcessor
+    params:
+      split_by: word
+      split_length: 1000
+  - name: FileTypeClassifier
+    type: FileTypeClassifier
+
+pipelines:
+  - name: query 
+    type: Query
+    nodes:
+      - name: Retriever
+        inputs: [Query]
+      - name: Ranker
+        inputs: [Retriever]
+  - name: indexing
+    type: Indexing
+    nodes:
+      - name: FileTypeClassifier
+        inputs: [File]
+      - name: TextFileConverter
+        inputs: [FileTypeClassifier.output_1]
+      - name: PDFFileConverter
+        inputs: [FileTypeClassifier.output_2]
+      - name: DocxFileConverter
+        inputs: [FileTypeClassifier.output_4]
+      - name: ImageFileConverter
+        inputs: [FileTypeClassifier.output_6]
+      - name: Preprocessor
+        inputs: [PDFFileConverter, TextFileConverter, DocxFileConverter, ImageFileConverter]
+      - name: Retriever
+        inputs: [Preprocessor]
+      - name: DocumentStore
+        inputs: [Retriever]
diff --git a/pipelines/utils/offline_ann.py b/pipelines/utils/offline_ann.py
@@ -82,7 +82,13 @@
 parser.add_argument(
     '--delete_index',
     action='store_true',
-    help='whether to delete existing index while updating index')
+    help='Whether to delete existing index while updating index')
+
+parser.add_argument(
+    '--share_parameters',
+    action='store_true',
+    help='Use to control the query and title models sharing the same parameters'
+)
 
 args = parser.parse_args()
 
@@ -126,6 +132,7 @@ def offline_ann(index_name, doc_dir):
         passage_embedding_model=args.passage_embedding_model,
         params_path=args.params_path,
         output_emb_size=args.embedding_dim,
+        share_parameters=args.share_parameters,
         max_seq_len_query=64,
         max_seq_len_passage=256,
         batch_size=16,
