Merge branch 'init_SentenceTransformer' into 'advana-release/v2.0.1'

Init sentence transformer

See merge request advana/gamechanger/gamechanger-ml-source!180

Description:

How to test:

Remember to attach the issue if it's not

gamechangerml/api/fastapi/model_loader.py

@@ -16,7 +16,8 @@
  from gamechangerml.src.search.sent_transformer.model import (
  SentenceSearcher,
  SentenceEncoder,
- SemanticSearcher
+ SemanticSearcher,
+ GcSentenceTransformer
  )
  from gamechangerml.src.search.doc_compare import (
  DocCompareSentenceEncoder,
@@ -68,6 +69,7 @@ def __init__(self):
  __document_compare_searcher = None
  __document_compare_encoder = None
  __semantic_searcher = None
+ __gc_sentence_transformer = None
 
  # Get methods for the models. If they don't exist try initializing them.
  def getQA(self):
@@ -161,6 +163,14 @@ def getSemanticSearcher(self):
  ModelLoader.initSemanticSearcher()
  return ModelLoader.__semantic_searcher
 
+ def getGcSentenceTransformer(self):
+ if ModelLoader.__gc_sentence_transformer == None:
+ logger.warning(
+ "semantic_searcher was not set and was attempted to be used. Running init"
+ )
+ ModelLoader.initGcSentenceTransformer()
+ return ModelLoader.__gc_sentence_transformer
+
  def set_error(self):
  logger.error("Models cannot be directly set. Must use init methods.")
 
@@ -178,6 +188,7 @@ def set_error(self):
  document_compare_searcher = property(getDocumentCompareSearcher, set_error)
  document_compare_encoder = property(getDocumentCompareEncoder, set_error)
  semantic_searcher = property(getSemanticSearcher, set_error)
+ gc_sentence_transformer = property(getGcSentenceTransformer, set_error)
 
  @staticmethod
  def initQA(qa_model_name=QA_MODEL.value):
@@ -385,6 +396,28 @@ def initSemanticSearcher(
  except Exception as e:
  logger.warning("** Could not load Similarity model for Semantic Searcher")
  logger.warning(e)
+
+ @staticmethod
+ def initGcSentenceTransformer(transformer_path=LOCAL_TRANSFORMERS_DIR.value):
+ """
+ initGcSentenceTransformer - loads Sentence Transformer on start
+ Args:
+ Returns:
+ """
+ logger.info(f"Loading GC sentence transformer model")
+ try:
+ ModelLoader.__gc_sentence_transformer = GcSentenceTransformer(
+ encoder_model_name=EmbedderConfig.BASE_MODEL,
+ transformer_path=transformer_path
+ )
+ encoder_model = ModelLoader.__gc_sentence_transformer.encoder_model
+ # set cache variable defined in settings.py
+ latest_intel_model_encoder.value = encoder_model
+ logger.info(f"** Loaded GC Sentence Transformer Model from {encoder_model}")
+
+ except Exception as e:
+ logger.warning("** Could not load Encoder model")
+ logger.warning(e)
 
  @staticmethod
  def initSparse(model_name=latest_intel_model_trans.value):

gamechangerml/api/fastapi/routers/search.py

@@ -93,9 +93,10 @@ async def semantic_search(
  Returns:
  results: list of floats
  """
+ logger.info(f"Hit embedSemanticQuery with json: {body}")
  query_text = body["query"]
- embeddings = MODELS.semantic_searcher.embed_query(query_text)
- return  list(embeddings.astype(float))
+ embeddings = MODELS.gc_sentence_transformer.embed_query(query_text)
+ return list(embeddings.astype(float))
 
 
 @router.post("/semanticSearch", status_code=200)

gamechangerml/api/fastapi/routers/startup.py

@@ -39,7 +39,8 @@
  MODELS.initTopics,
  MODELS.initRecommender,
  MODELS.initDocumentCompareSearcher,
- MODELS.initSemanticSearcher
+ MODELS.initSemanticSearcher,
+ MODELS.initGcSentenceTransformer
 ]
 
 

gamechangerml/src/search/sent_transformer/model.py

@@ -1,6 +1,7 @@
 from txtai.embeddings import Embeddings
 from txtai.pipeline import Similarity
 from txtai.ann import ANN
+from sentence_transformers import SentenceTransformer
 
 import os
 import numpy as np
@@ -474,4 +475,21 @@ def search(
 
  def embed_query(self, query_text):
  embeddings = self.embedder.transform(query_text)
+ return embeddings
+
+class GcSentenceTransformer(object):
+ """
+ Instantiates Sentence Transformer model 
+
+ Args:
+ transformer_path (str): Path to transformer directory
+ encoder_model (str): Model name supported by sentence_transformers
+ """
+
+ def __init__(self, encoder_model_name, transformer_path):
+ self.encoder_model = SentenceTransformer(os.path.join(transformer_path, encoder_model_name))
+
+ def embed_query(self, query_text):
+ query_text = " ".join(preprocess(query_text))
+ embeddings = self.encoder_model.encode(query_text)
  return embeddings