EFSR-885 gender age plugins (#281)

* EFRS-885 Add age/gender recognition as an independent plugin

embedding-calculator/Dockerfile

@@ -13,8 +13,6 @@ WORKDIR /app/ml
 COPY requirements.txt .
 RUN pip --no-cache-dir install -r requirements.txt
 
-ARG SCANNER=Facenet2018
-ENV SCANNER=$SCANNER
 ARG BE_VERSION
 ARG APP_VERSION_STRING
 ENV BE_VERSION=$BE_VERSION
@@ -26,35 +24,25 @@ ENV PYTHONUNBUFFERED=0
 ENV JOBLIB_MULTIPROCESSING=0
 
 # download ML models
-ARG DETECTION_MODEL
-ARG CALCULATION_MODEL
-ENV DETECTION_MODEL=${DETECTION_MODEL:-retinaface_r50_v1}
-ENV CALCULATION_MODEL=${CALCULATION_MODEL:-arcface_r100_v1}
+ARG INTEL_OPTIMIZATION=false
+ARG GPU_IDX=-1
+ENV GPU_IDX=$GPU_IDX INTEL_OPTIMIZATION=$INTEL_OPTIMIZATION
+ARG FACE_DETECTION_PLUGIN="facenet.FaceDetector"
+ARG CALCULATION_PLUGIN="facenet.Calculator"
+ARG EXTRA_PLUGINS=""
+ENV FACE_DETECTION_PLUGIN=$FACE_DETECTION_PLUGIN CALCULATION_PLUGIN=$CALCULATION_PLUGIN \
+    EXTRA_PLUGINS=$EXTRA_PLUGINS
+COPY src src
 COPY srcext srcext
-COPY pytest.ini .
-COPY *.sh ./
-RUN chmod +x *.sh
-RUN ./prepare_scanners.sh
-
-# install InsightFace packages
-ARG INTEL_OPTIMIZATION
-ARG GPU_IDX
-ENV GPU_IDX=$GPU_IDX
-ENV MXNET_MOD=${GPU_IDX:+cu101}${INTEL_OPTIMIZATION:+mkl}
-ENV MXNET_LIB=mxnet${MXNET_MOD:+-$MXNET_MOD}
-ENV MXNET_VER="<1.7"
-
-RUN if [[ "$SCANNER" == "InsightFace" ]]; then \
-      pip --no-cache-dir install "$MXNET_LIB$MXNET_VER" -e srcext/insightface/python-package; \
-    fi
+RUN python -m src.services.facescan.plugins.setup
 
 # copy rest of the code
-COPY src src
 COPY tools tools
 COPY sample_images sample_images
 
 # run tests
 ARG SKIP_TESTS
+COPY pytest.ini .
 RUN if [ -z $SKIP_TESTS  ]; then pytest -m "not performance" /app/ml/src; fi
 
 EXPOSE 3000

embedding-calculator/gpu.Dockerfile

@@ -12,6 +12,7 @@ ARG CUDNN_MAJOR_VERSION=7
 ARG LIB_DIR_PREFIX=x86_64
 ARG LIBNVINFER=6.0.1-1
 ARG LIBNVINFER_MAJOR_VERSION=6
+ENV CUDA=$CUDA
 
 # Needed for string substitution
 SHELL ["/bin/bash", "-c"]
@@ -66,7 +67,7 @@ RUN ln -s $(which $PYTHON) /usr/local/bin/python
 
 
 # Variables for MXNET
-ENV MXNET=mxnet_cu101mkl MXNET_CPU_WORKER_NTHREADS=24
+ENV MXNET_CPU_WORKER_NTHREADS=24
 ENV MXNET_ENGINE_TYPE=ThreadedEnginePerDevice MXNET_CUDNN_AUTOTUNE_DEFAULT=0
 
 # No access to GPU devices in the build stage, so skip tests

embedding-calculator/requirements.txt

@@ -3,6 +3,7 @@ cached-property==1.5.2
 colour==0.1.5
 flasgger==0.9.5
 Flask==1.1.2
+gdown~=3.12
 Werkzeug==1.0.1
 
 # tests

embedding-calculator/src/_endpoints.py

@@ -11,48 +11,59 @@
 #  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
 #  or implied. See the License for the specific language governing
 #  permissions and limitations under the License.
-from typing import List
+from typing import List, Optional
 
 from flask import request
 from flask.json import jsonify
 from werkzeug.exceptions import BadRequest
 
 from src.constants import ENV
 from src.exceptions import NoFaceFoundError
+from src.services.facescan.plugins import managers
 from src.services.facescan.scanner.facescanners import scanner
 from src.services.flask_.constants import ARG
 from src.services.flask_.needs_attached_file import needs_attached_file
 from src.services.imgtools.read_img import read_img
+from src.services.utils.pyutils import Constants
 
 
 def endpoints(app):
     @app.route('/status')
     def status_get():
+        available_plugins = {p.slug: str(p)
+                             for p in managers.plugin_manager.plugins}
+        calculator = managers.plugin_manager.calculator
         return jsonify(status='OK', build_version=ENV.BUILD_VERSION,
-                       calculator_version=ENV.SCANNER)
+                       calculator_version=str(calculator),
+                       available_plugins=available_plugins)
 
     @app.route('/find_faces', methods=['POST'])
     @needs_attached_file
     def find_faces_post():
-        faces = scanner.find_faces(
+        detector = managers.plugin_manager.detector
+        face_plugins = managers.plugin_manager.filter_face_plugins(
+            _get_face_plugin_names()
+        )
+        faces = detector(
             img=read_img(request.files['file']),
-            det_prob_threshold=_get_det_prob_threshold(request),
+            det_prob_threshold=_get_det_prob_threshold(),
+            face_plugins=face_plugins
         )
-        faces = _limit(faces, request.values.get(ARG.LIMIT))
-        return jsonify(calculator_version=scanner.ID, result=faces)
+        plugins_versions = {p.slug: str(p) for p in [detector] + face_plugins}
+        return jsonify(results=faces, plugins_versions=plugins_versions)
 
     @app.route('/scan_faces', methods=['POST'])
     @needs_attached_file
     def scan_faces_post():
         faces = scanner.scan(
             img=read_img(request.files['file']),
-            det_prob_threshold=_get_det_prob_threshold(request)
+            det_prob_threshold=_get_det_prob_threshold()
         )
         faces = _limit(faces, request.values.get(ARG.LIMIT))
         return jsonify(calculator_version=scanner.ID, result=faces)
 
 
-def _get_det_prob_threshold(request):
+def _get_det_prob_threshold():
     det_prob_threshold_val = request.values.get(ARG.DET_PROB_THRESHOLD)
     if det_prob_threshold_val is None:
         return None
@@ -62,6 +73,14 @@ def _get_det_prob_threshold(request):
     return det_prob_threshold
 
 
+def _get_face_plugin_names() -> Optional[List[str]]:
+    if ARG.FACE_PLUGINS not in request.values:
+        return
+    return [
+        name for name in Constants.split(request.values[ARG.FACE_PLUGINS])
+    ]
+
+
 def _limit(faces: List, limit: str = None) -> List:
     """
     >>> _limit([1, 2, 3], None)

embedding-calculator/src/constants.py

@@ -14,24 +14,25 @@
 
 import logging
 
-from src.services.utils.pyutils import get_env, Constants
+from src.services.utils.pyutils import get_env, get_env_split, get_env_bool, Constants
 
 _DEFAULT_SCANNER = 'Facenet2018'
 
 
 class ENV(Constants):
     ML_PORT = int(get_env('ML_PORT', '3000'))
-    SCANNER = get_env('SCANNER', _DEFAULT_SCANNER)
-    SCANNERS = [SCANNER]
     IMG_LENGTH_LIMIT = int(get_env('IMG_LENGTH_LIMIT', '640'))
 
+    FACE_DETECTION_PLUGIN = get_env('FACE_DETECTION_PLUGIN', 'facenet.FaceDetector')
+    CALCULATION_PLUGIN = get_env('CALCULATION_PLUGIN', 'facenet.Calculator')
+    EXTRA_PLUGINS = get_env_split('EXTRA_PLUGINS', '')
+
     LOGGING_LEVEL_NAME = get_env('LOGGING_LEVEL_NAME', 'debug').upper()
     IS_DEV_ENV = get_env('FLASK_ENV', 'production') == 'development'
     BUILD_VERSION = get_env('APP_VERSION_STRING', 'dev')
 
     GPU_IDX = int(get_env('GPU_IDX', '-1'))
-    DETECTION_MODEL = get_env('DETECTION_MODEL', 'retinaface_r50_v1')
-    CALCULATION_MODEL = get_env('CALCULATION_MODEL', 'arcface_r100_v1')
+    INTEL_OPTIMIZATION = get_env_bool('INTEL_OPTIMIZATION')
 
 
 LOGGING_LEVEL = logging._nameToLevel[ENV.LOGGING_LEVEL_NAME]

embedding-calculator/src/docs/find_faces_post.yml

@@ -2,7 +2,7 @@ tags:
   - Core
 summary: 'Find faces in the given image and return their bounding boxes.'
 description: 'Returns bounding boxes of detected faces on the image.'
-operationId: scanFacesPost
+operationId: findFacesPost
 consumes:
   - multipart/form-data
 produces:

embedding-calculator/src/exceptions.py

@@ -37,20 +37,10 @@ class OneDimensionalImageIsGivenError(BadRequest):
     description = "Given image has only one dimension"
 
 
-class MoreThanOneFaceFoundError(BadRequest):
-    description = "Found more than one face in the given image"
-
-
 class ClassifierIsAlreadyTrainingError(Locked):
     description = "Classifier training is already in progress"
 
 
-class NoTrainedEmbeddingClassifierFoundError(BadRequest):
-    description = "No classifier model is yet trained, please train a classifier first. If the problem persists, " \
-                  "check the amount of unique faces saved, and whether all face embeddings have been migrated to " \
-                  f"version '{ENV.SCANNER}'"
-
-
 class NoFileFoundInDatabaseError(InternalServerError):
     description = "File is not found in the database"
 

embedding-calculator/src/services/dto/plugin_result.py

@@ -0,0 +1,50 @@
+import attr
+from typing import Tuple, List, Optional, Dict
+
+from src.services.dto.bounding_box import BoundingBoxDTO
+from src.services.dto.json_encodable import JSONEncodable
+from src.services.imgtools.types import Array1D, Array3D
+
+
+class PluginResultDTO(JSONEncodable):
+    def to_json(self) -> dict:
+        """ Serialize only public properties """
+        return {k: v for k, v in self.__dict__.items() if not k.startswith('_')}
+
+
+@attr.s(auto_attribs=True, frozen=True)
+class EmbeddingDTO(PluginResultDTO):
+    embedding: Array1D
+
+
+@attr.s(auto_attribs=True, frozen=True)
+class GenderDTO(PluginResultDTO):
+    gender: str
+    gender_probability: float = attr.ib(converter=float, default=None)
+
+
+@attr.s(auto_attribs=True, frozen=True)
+class AgeDTO(PluginResultDTO):
+    age: Tuple[int, int]
+    age_probability: float = attr.ib(converter=float, default=None)
+
+
+@attr.s(auto_attribs=True)
+class FaceDTO(PluginResultDTO):
+    box: BoundingBoxDTO
+    _img: Optional[Array3D]
+    _face_img: Optional[Array3D]
+    _plugins_dto: List[PluginResultDTO] = attr.Factory(list)
+    execution_time: Dict[str, float] = attr.Factory(dict)
+
+    def to_json(self):
+        data = super().to_json()
+        for plugin_dto in self._plugins_dto:
+            data.update(plugin_dto.to_json())
+        return data
+
+    @property
+    def embedding(self):
+        for dto in self._plugins_dto:
+            if isinstance(dto, EmbeddingDTO):
+                return dto.embedding

embedding-calculator/src/services/dto/scanned_face.py

@@ -12,7 +12,7 @@
 #  or implied. See the License for the specific language governing
 #  permissions and limitations under the License.
 
-from typing import Union
+from typing import Union, Optional
 
 import attr
 
@@ -22,23 +22,27 @@
 from src.services.imgtools.types import Array1D, Array3D
 
 
+@attr.s(auto_attribs=True)
+class Face(JSONEncodable):
+    _img: Optional[Array3D]
+    _face_img: Optional[Array3D]
+    box: BoundingBoxDTO
+
+
 @attr.s(auto_attribs=True, frozen=True)
 class ScannedFaceDTO(JSONEncodable):
     box: BoundingBoxDTO
     embedding: Array1D
 
 
-class ScannedFace(JSONEncodable):
-    def __init__(self, box: BoundingBoxDTO, embedding: Array1D, img: Union[Array3D, None], face_img: Array3D = None):
-        self.box = box
-        self.embedding = embedding
-        self.img = img
-        self._face_img = face_img
+@attr.s(auto_attribs=True)
+class ScannedFace(Face):
+    embedding: Array1D
 
     @property
     def face_img(self):
         if not self._face_img:
-            self._face_img = crop_img(self.img, self.box)
+            self._face_img = crop_img(self._img, self.box)
         return self._face_img
 
     @property
@@ -53,4 +57,5 @@ def from_request(cls, result):
                                               y_min=box_result['y_min'],
                                               y_max=box_result['y_max'],
                                               probability=box_result['probability']),
-                           embedding=result['embedding'], img=None)
+                           embedding=result['embedding'],
+                           img=None, face_img=None)