fixed multiprocessing and optimized model preloading

dev/docker-compose.yml

@@ -71,6 +71,7 @@ services:
       - MAX_REQUEST_SIZE=${max_request_size}B
       - CONNECTION_TIMEOUT=${connection_timeout:-10000}
       - READ_TIMEOUT=${read_timeout:-60000}
+      - DISTANCE_TYPE=${distance_type:-custom}
 
   compreface-fe:
     image: ${registry}compreface-fe:${FE_VERSION}
@@ -100,12 +101,13 @@ services:
     environment:
       - ML_PORT=3000
       - IMG_LENGTH_LIMIT=${max_detect_size}
-      #- UWSGI_PROCESSES=${uwsgi_processes:-1}
-      #- UWSGI_THREADS=${uwsgi_threads:-1}
+      - UWSGI_PROCESSES=${uwsgi_processes:-1}
+      - UWSGI_THREADS=${uwsgi_threads:-1}
       - CUDA_VISIBLE_DEVICES=0  ##
       - NVIDIA_VISIBLE_DEVICES=0
       - TZ=utc
       - DEBIAN_FRONTEND=noninteractive
+      - DEVICE=cuda:0
     healthcheck:
       test: curl --fail http://localhost:3000/healthcheck || exit 1
       interval: 10s

embedding-calculator/src/services/facescan/plugins/adaface/inference.py

@@ -15,15 +15,21 @@
     model_mame:model_path,
 }
 
+preloaded = ""
+
 def load_pretrained_model(architecture=model_mame):
-    # load model and pretrained statedict
-    assert architecture in adaface_models.keys()
-    model = net.build_model(architecture)
-    statedict = torch.load(adaface_models[architecture],map_location =device)['state_dict']
-    model_statedict = {key[6:]:val for key, val in statedict.items() if key.startswith('model.')}
-    model.load_state_dict(model_statedict)
-    model.eval()
-    return model
+    global preloaded
+
+    if preloaded == "":
+        # load model and pretrained statedict
+        assert architecture in adaface_models.keys()
+        preloaded = net.build_model(architecture)
+        print("load adaface model")
+        statedict = torch.load(adaface_models[architecture],map_location =device)['state_dict']
+        model_statedict = {key[6:]:val for key, val in statedict.items() if key.startswith('model.')}
+        preloaded.load_state_dict(model_statedict)
+        preloaded.eval()
+    return preloaded
 
 def to_input(pil_rgb_image):
     np_img = np.array(pil_rgb_image)
@@ -53,6 +59,7 @@ def inference_detector(image_path):
 
         detected["result"].append(face)
 
+    print(f'Processed {len(detected["result"])} faces')
     return detected
 
 def inference_scaner(image_path):

embedding-calculator/src/services/facescan/plugins/pytorch_detector/detect.py

@@ -355,20 +355,34 @@ def warp_and_crop_face(src_img,
 
     return face_img
 
+preloaded = ""
+
+def preload_model(device, cfg):
+    global preloaded
+
+    if preloaded == "":
+        print("load RetinaFace model")
+        preloaded = RetinaFace(cfg=cfg, phase = 'test')
+        preloaded = load_model(preloaded, args.trained_model, args.cpu)
+        preloaded.eval()
+        cudnn.benchmark = True
+
+        preloaded = preloaded.to(device)
+    return preloaded
+
 def retina_detector(image_pil, content_type):
     torch.set_grad_enabled(False)
+
     cfg = None
     if args.network == "mobile0.25":
         cfg = cfg_mnet
     elif args.network == "resnet50":
         cfg = cfg_re50
-    # net and model
-    net = RetinaFace(cfg=cfg, phase = 'test')
-    net = load_model(net, args.trained_model, args.cpu)
-    net.eval()
-    cudnn.benchmark = True
+
     device = torch.device("cpu" if args.cpu else "cuda")
-    net = net.to(device)
+    # net and model
+
+    net = preload_model(device, cfg)
 
     resize = 1
 

embedding-calculator/uwsgi.ini

@@ -2,7 +2,8 @@
 module = src.app:wsgi_app()
 uid = www-data
 gid = www-data
-master = true
+#master = true
+lazy=true
 http-socket = 0.0.0.0:3000
 vacuum = true
 die-on-term = true