Adjust Docker and REST API to allow TransformsReader Class (#180)

Dockerfile

@@ -14,9 +14,12 @@ COPY rest_api /home/user/rest_api
 # copy saved FARM models
 COPY models /home/user/models
 
-# copy sqlite db if needed for testing
+# optional : copy sqlite db if needed for testing
 #COPY qa.db /home/user/
 
+# optional: copy data directory containing docs for indexing
+#COPY data /home/user/data
+
 EXPOSE 8000
 
 # cmd for running the API

Dockerfile-GPU

@@ -22,6 +22,9 @@ COPY models /home/user/models
 # Optional: copy sqlite db if needed for testing
 #COPY qa.db /home/user/
 
+# Optional: copy data directory containing docs for indexing
+#COPY data /home/user/data
+
 EXPOSE 8000
 
 ENV LC_ALL=C.UTF-8

docker-compose.yml

@@ -1,19 +1,36 @@
 version: '3'
 services:
   haystack-api:
-    image: "deepset/haystack-cpu:0.2.0"
+    build:
+      context: .
+      dockerfile: Dockerfile
+    image: "deepset/haystack-cpu:latest"
     ports:
       - 8000:8000
+    volumes:
+      # Optional: mount your own models from disk into the container
+      - "./models:/home/user/models"
     environment:
-      # see haystack/api/config.py for additional variables to configure.
-      # load reader model from transformers' model hub.
-      - READER_MODEL_PATH=deepset/roberta-base-squad2
+      # See rest_api/config.py for more variables that you can configure here.
       - DB_HOST=elasticsearch
+      - USE_GPU=False
+      # Load a model from transformers' model hub or a local path into the FARMReader.
+      - READER_MODEL_PATH=deepset/roberta-base-squad2
+      # - READER_MODEL_PATH=home/user/models/roberta-base-squad2
+      # Alternative: If you want to use the TransformersReader (e.g. for loading a local model in transformers format):
+      # - READER_USE_TRANSFORMERS=True
+      # - READER_MODEL_PATH=/home/user/models/roberta-base-squad2
+      # - READER_TOKENIZER=/home/user/models/roberta-base-squad2
     restart: always
-
+    depends_on:
+      - elasticsearch
+    command: "/bin/bash -c 'sleep 15 && gunicorn rest_api.application:app -b 0.0.0.0 -k uvicorn.workers.UvicornWorker --workers 1 --timeout 180'"
   elasticsearch:
-    # This demo image contains Game of Thrones Wikipedia articles indexed.
-    # For starting a new Elasticsearch instance, replace image with "elasticsearch:7.6.1".
-    image: "deepset/elasticsearch-game-of-thrones"
+    # This will start an empty elasticsearch instance (so you have to add your documents yourself)
+    image: "elasticsearch:7.6.1"
+    # If you want a demo image instead that is "ready-to-query" with some indexed Game of Thrones articles:
+    # image: "deepset/elasticsearch-game-of-thrones"
+    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", "False").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

@@ -13,10 +13,11 @@
     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.sparse import ElasticsearchRetriever
 from haystack.retriever.dense import EmbeddingRetriever
@@ -54,17 +55,26 @@
     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