Docker available for TransformsReader Class

Dockerfile

@@ -16,6 +16,7 @@ COPY models /home/user/models
 
 # copy sqlite db if needed for testing
 #COPY qa.db /home/user/
+COPY data /home/user/data
 
 EXPOSE 8000
 

docker-compose-custom.yml

@@ -0,0 +1,28 @@
+version: '3'
+services:
+  haystack-api:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    ports:
+      - 8000:8000
+    environment:
+      # see rest_api/config.py for additional variables to configure.
+      - READER_MODEL_PATH=distilbert-base-uncased-distilled-squad
+      # - READER_MODEL_PATH=fmikaelian/camembert-base-fquad
+      # - READER_MODEL_PATH=deepset/roberta-base-squad2
+      # - READER_MODEL_PATH=/home/user/models/roberta-base-squad2
+      - READER_TOKENIZER=distilbert-base-uncased
+      - READER_USE_TRANSFORMERS=True
+      - DB_HOST=elasticsearch
+      - USE_GPU=False
+    restart: always
+    depends_on:
+      - elasticsearch
+
+  elasticsearch:
+    image: "elasticsearch:7.6.1"
+    ports:
+      - 9200:9200
+    environment:
+      - discovery.type=single-node

haystack/reader/transformers.py

@@ -37,7 +37,7 @@ def __init__(
         :param tokenizer: name of the tokenizer (usually the same as model)
         :param context_window_size: num of chars (before and after the answer) to return as "context" for each answer.
                             The context usually helps users to understand if the answer really makes sense.
-        :param use_gpu: < 1  -> use cpu
+        :param use_gpu: < 0  -> use cpu
                         >= 0 -> ordinal of the gpu to use
         """
         self.model = pipeline("question-answering", model=model, tokenizer=tokenizer, device=use_gpu)

rest_api/config.py

@@ -6,6 +6,7 @@
 
 # Resources / Computation
 USE_GPU = os.getenv("USE_GPU", "True").lower() == "true"
+GPU_NUMBER = int(os.getenv("GPU_NUMBER", 1))
 MAX_PROCESSES = int(os.getenv("MAX_PROCESSES", 4))
 BATCHSIZE = int(os.getenv("BATCHSIZE", 50))
 CONCURRENT_REQUEST_PER_WORKER = int(os.getenv("CONCURRENT_REQUEST_PER_WORKER", 4))
@@ -26,6 +27,8 @@
 
 # Reader
 READER_MODEL_PATH = os.getenv("READER_MODEL_PATH", None)
+READER_USE_TRANSFORMERS = os.getenv("READER_USE_TRANSFORMERS", "True").lower() == "true"
+READER_TOKENIZER = os.getenv("READER_TOKENIZER", None)
 CONTEXT_WINDOW_SIZE = int(os.getenv("CONTEXT_WINDOW_SIZE", 500))
 DEFAULT_TOP_K_READER = int(os.getenv("DEFAULT_TOP_K_READER", 5))
 TOP_K_PER_CANDIDATE = int(os.getenv("TOP_K_PER_CANDIDATE", 3))

rest_api/controller/search.py

@@ -12,10 +12,11 @@
     EMBEDDING_DIM, EMBEDDING_FIELD_NAME, EXCLUDE_META_DATA_FIELDS, EMBEDDING_MODEL_PATH, USE_GPU, READER_MODEL_PATH, \
     BATCHSIZE, CONTEXT_WINDOW_SIZE, TOP_K_PER_CANDIDATE, NO_ANS_BOOST, MAX_PROCESSES, MAX_SEQ_LEN, DOC_STRIDE, \
     DEFAULT_TOP_K_READER, DEFAULT_TOP_K_RETRIEVER, CONCURRENT_REQUEST_PER_WORKER, FAQ_QUESTION_FIELD_NAME, \
-    EMBEDDING_MODEL_FORMAT
+    EMBEDDING_MODEL_FORMAT, READER_USE_TRANSFORMERS, READER_TOKENIZER, GPU_NUMBER
 from rest_api.controller.utils import RequestLimiter
 from haystack.database.elasticsearch import ElasticsearchDocumentStore
 from haystack.reader.farm import FARMReader
+from haystack.reader.transformers import TransformersReader
 from haystack.retriever.base import BaseRetriever
 from haystack.retriever.elasticsearch import ElasticsearchRetriever, EmbeddingRetriever
 
@@ -52,17 +53,26 @@
     retriever = ElasticsearchRetriever(document_store=document_store)
 
 if READER_MODEL_PATH:  # for extractive doc-qa
-    reader = FARMReader(
-        model_name_or_path=str(READER_MODEL_PATH),
-        batch_size=BATCHSIZE,
-        use_gpu=USE_GPU,
-        context_window_size=CONTEXT_WINDOW_SIZE,
-        top_k_per_candidate=TOP_K_PER_CANDIDATE,
-        no_ans_boost=NO_ANS_BOOST,
-        num_processes=MAX_PROCESSES,
-        max_seq_len=MAX_SEQ_LEN,
-        doc_stride=DOC_STRIDE,
-    )  # type: Optional[FARMReader]
+    if READER_USE_TRANSFORMERS:
+        use_gpu = -1 if not USE_GPU else GPU_NUMBER
+        reader = TransformersReader(
+            model=str(READER_MODEL_PATH),
+            use_gpu=use_gpu,
+            context_window_size=CONTEXT_WINDOW_SIZE,
+            tokenizer=str(READER_TOKENIZER)
+        )  # type: Optional[FARMReader]
+    else:
+        reader = FARMReader(
+            model_name_or_path=str(READER_MODEL_PATH),
+            batch_size=BATCHSIZE,
+            use_gpu=USE_GPU,
+            context_window_size=CONTEXT_WINDOW_SIZE,
+            top_k_per_candidate=TOP_K_PER_CANDIDATE,
+            no_ans_boost=NO_ANS_BOOST,
+            num_processes=MAX_PROCESSES,
+            max_seq_len=MAX_SEQ_LEN,
+            doc_stride=DOC_STRIDE,
+        )  # type: Optional[FARMReader]
 else:
     reader = None  # don't need one for pure FAQ matching