Fix colliding cuda context

- Extract face_recognition and dlib into its own context
- Prevents initilizing the cuda context multiple times from dlib and pytorch

api/face_recognition.py

@@ -0,0 +1,25 @@
+import requests
+import numpy as np
+
+
+def get_face_encodings(image_path, known_face_locations):
+    json = {
+        "source": image_path,
+        "face_locations": known_face_locations,
+    }
+    face_encoding = requests.post(
+        "http://localhost:8005/face-encodings", json=json
+    ).json()
+
+    face_encodings_list = face_encoding["encodings"]
+    face_encodings = [np.array(enc) for enc in face_encodings_list]
+
+    return face_encodings
+
+
+def get_face_locations(image_path, model="hog"):
+    json = {"source": image_path, "model": model}
+    face_locations = requests.post(
+        "http://localhost:8005/face-locations", json=json
+    ).json()
+    return face_locations["face_locations"]

api/models/photo.py

@@ -12,6 +12,7 @@
 from django.db.models import Q
 from django.db.utils import IntegrityError
 from api.im2txt.sample import Im2txt
+from api.face_recognition import get_face_encodings, get_face_locations
 
 import api.date_time_extractor as date_time_extractor
 import api.models
@@ -702,20 +703,18 @@ def _extract_faces(self, second_try=False):
                     ).count()
                     image_path = self.image_hash + "_" + str(idx_face) + ".jpg"
 
-                    import face_recognition
-
-                    face_encodings = face_recognition.face_encodings(
-                        image, known_face_locations=[(top, right, bottom, left)]
-                    )
-
+                    face_encoding = get_face_encodings(
+                        self.thumbnail_big.path,
+                        known_face_locations=[(top, right, bottom, left)],
+                    )[0]
                     face = api.models.face.Face(
                         image_path=image_path,
                         photo=self,
                         location_top=top,
                         location_right=right,
                         location_bottom=bottom,
                         location_left=left,
-                        encoding=face_encodings[0].tobytes().hex(),
+                        encoding=face_encoding.tobytes().hex(),
                         person=person,
                         cluster=unknown_cluster,
                     )
@@ -729,16 +728,14 @@ def _extract_faces(self, second_try=False):
                     logger.debug(f"Created face {face} from {self.main_file.path}")
             return
 
-        import face_recognition
-
         try:
             image = np.array(PIL.Image.open(self.thumbnail_big.path))
 
             face_locations = []
             # Create
             try:
-                face_locations = face_recognition.face_locations(
-                    image, model=self.owner.face_recognition_model.lower()
+                face_locations = get_face_locations(
+                    self.thumbnail_big.path, model=self.owner.face_recognition_model.lower()
                 )
             except Exception as e:
                 logger.info(
@@ -747,8 +744,8 @@ def _extract_faces(self, second_try=False):
                 logger.info(e)
 
             if len(face_locations) > 0:
-                face_encodings = face_recognition.face_encodings(
-                    image, known_face_locations=face_locations
+                face_encodings = get_face_encodings(
+                    self.thumbnail_big.path, known_face_locations=face_locations
                 )
                 for idx_face, face in enumerate(zip(face_encodings, face_locations)):
                     face_encoding = face[0]

service/face_recognition/main.py

@@ -0,0 +1,55 @@
+import gevent
+from flask import Flask, request
+from gevent.pywsgi import WSGIServer
+import face_recognition
+import numpy as np
+import PIL
+
+app = Flask(__name__)
+
+
+def log(message):
+    print("face_recognition: {}".format(message))
+
+
+@app.route("/face-encodings", methods=["POST"])
+def create_face_encodings():
+    try:
+        data = request.get_json()
+        source = data["source"]
+        face_locations = data["face_locations"]
+    except Exception:
+        return "", 400
+
+    image = np.array(PIL.Image.open(source))
+    face_encodings = face_recognition.face_encodings(
+        image,
+        known_face_locations=face_locations,
+    )
+    # Convert NumPy arrays to Python lists
+    face_encodings_list = [enc.tolist() for enc in face_encodings]
+    # Log number of face encodings
+    log(f"created face_encodings={len(face_encodings_list)}")
+    return {"encodings": face_encodings_list}, 201
+
+
+@app.route("/face-locations", methods=["POST"])
+def create_face_locations():
+    try:
+        data = request.get_json()
+        source = data["source"]
+        model = data["model"]
+    except Exception:
+        return "", 400
+
+    image = np.array(PIL.Image.open(source))
+    face_locations = face_recognition.face_locations(image, model=model)
+    log(f"created face_location={face_locations}")
+    return {"face_locations": face_locations}, 201
+
+
+if __name__ == "__main__":
+    log("service starting")
+    server = WSGIServer(("0.0.0.0", 8005), app)
+    server_thread = gevent.spawn(server.serve_forever)
+    gevent.joinall([server_thread])