Replace check by early return

.github/workflows/pytest.yml

@@ -16,31 +16,32 @@ jobs:
         run: |
           docker run -d -it --rm nvcr.io/nvidia/pytorch:23.07-py3 echo "I have to be startet to authenticate against nvcr.io"
       - name: build
-        if: github.ref != 'refs/heads/release'
         run: |
           cd trainer
           ./docker.sh b test_latest
-      - name: buildnocaches
-        if: github.ref == 'refs/heads/release'
-        run: |
-          cd trainer
-          ./docker.sh bnc test_latest
       - name: run
         run: |
           docker run -d -it --rm -v /home/zauberzeug/data:/data -h n6 -e LOOP_HOST=$LOOP_HOST  -e LOOP_USERNAME=$LOOP_USERNAME -e LOOP_PASSWORD=$LOOP_PASSWORD -e YOLOV5_MODE=DETECTION --name yolov5_trainer_node_github_action --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all --gpus all --ipc host zauberzeug/yolov5-trainer:latest
       - name: tests
         run: |
           docker exec yolov5_trainer_node_github_action pytest -vv
-      - name: publish
-        if: github.ref == 'refs/heads/release'
-        run: |
-          cd trainer
-          ./docker.sh p test_latest
       - name: cleanup
         if: always() # also execute when pytest fails
         run: |
           docker stop yolov5_trainer_node_github_action | echo "Nothing to do. No container running."
 
+  pytest_detector_cpu:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v2
+      - name: build
+        run: |
+          cd detector_cpu
+          ./docker.sh b test_latest
+      - name: run tests
+        run: |
+          docker run --rm zauberzeug/yolov5-detector:latest-cpu pytest -vv
+
   slack:
     needs:
       - pytest

detector_cpu/cloud_cpu.dockerfile

@@ -28,12 +28,13 @@ RUN pip3 install IPython
 ARG NODE_LIB_VERSION
 RUN pip3 install "learning_loop_node==${NODE_LIB_VERSION}"
 
+ADD ./trainer/app_code/yolov5 /yolov5_node/detector_cpu/app_code/yolov5
+RUN pip install --no-cache -r /yolov5_node/detector_cpu/app_code/yolov5/requirements.txt
+
 ADD ./detector_cpu /yolov5_node/detector_cpu/
-ADD ./trainer/app_code /yolov5_node/detector_cpu/app_code
 RUN rm -f /yolov5_node/detector_cpu/.env
 RUN ln -sf /yolov5_node/detector_cpu /app
 
-RUN pip install --no-cache -r /yolov5_node/detector_cpu/app_code/yolov5/requirements.txt
 
 
 WORKDIR /app

detector_cpu/docker.sh

@@ -34,7 +34,11 @@ build_args=" --build-arg NODE_LIB_VERSION=$NODE_LIB_VERSION"
 
 
 dockerfile="cloud_cpu.dockerfile"
-image="zauberzeug/yolov5-detector:$SEMANTIC_VERSION-nlv$NODE_LIB_VERSION-cloud-cpu"
+if [ "$2" = "test_latest" ]; then
+    image="zauberzeug/yolov5-detector:latest-cpu"
+else
+    image="zauberzeug/yolov5-detector:$SEMANTIC_VERSION-nlv$NODE_LIB_VERSION-cloud-cpu"
+fi
 
 
 # ========================== RUN CONFIGURATION =========================================
@@ -61,7 +65,11 @@ fi
 # ========================== COMMAND EXECUTION =========================================
 
 cmd=$1
-cmd_args=${@:2}
+if [ "$2" = "test_latest" ]; then
+    cmd_args=${@:3}
+else
+    cmd_args=${@:2}
+fi
 set -x
 case $cmd in
     b | build)

detector_cpu/tests/test_yolov5_cpu.py

@@ -0,0 +1,46 @@
+import numpy as np
+
+from ..yolov5_detector import Yolov5Detector
+
+# pylint: disable=protected-access
+
+
+def test_postprocess_empty():
+    boxes, scores, class_id = Yolov5Detector()._post_process(np.empty(shape=(0, 8)), 100, 100, 0.2, 0.45)
+    assert len(boxes) == 0
+    assert len(scores) == 0
+    assert len(class_id) == 0
+
+
+def test_postprocess_conf_thresh_filtered_conf():
+    data = np.array([[0, 0, 10, 10, 0.1, 0.8, 0.8]])
+    boxes, scores, class_id = Yolov5Detector()._post_process(data, 100, 100, 0.2, 0.45)
+    assert len(boxes) == 0
+    assert len(scores) == 0
+    assert len(class_id) == 0
+
+
+def test_postprocess_conf_thresh_filtered_iou():
+    data = np.array(
+        [[0.5, 0, 0.1, 0.1, 0.95],
+         [0.5, 0, 0.1, 0.11, 0.9]]
+    )
+    detector = Yolov5Detector()
+    detector.input_size = 100
+    boxes, scores, class_id = detector._post_process(data, 100, 100, 0.2, 0.45)
+    assert len(boxes) == 1
+    assert len(scores) == 1
+    assert len(class_id) == 1
+
+
+def test_postprocess_conf_thresh_not_filtered():
+    data = np.array(
+        [[0.5, 0, 0.1, 0.1, 0.95],
+         [0, 0.5, 0.1, 0.1, 0.9]]
+    )
+    detector = Yolov5Detector()
+    detector.input_size = 100
+    boxes, scores, class_id = detector._post_process(data, 100, 100, 0.2, 0.45)
+    assert len(boxes) == 2
+    assert len(scores) == 2
+    assert len(class_id) == 2

detector_cpu/yolov5_detector.py

@@ -34,10 +34,8 @@ def init(self) -> None:
             raise RuntimeError("model_info.resolution must be an integer > 0")
 
         pt_file = f'{self.model_info.model_root_path}/model.pt'
-        yolov5_path = os.path.join(
-            os.path.dirname(__file__), 'app_code', 'yolov5')
-        self.yolov5 = torch.hub.load(
-            yolov5_path, 'custom', pt_file, source='local')
+        yolov5_path = os.path.join(os.path.dirname(__file__), 'app_code', 'yolov5')
+        self.yolov5 = torch.hub.load(yolov5_path, 'custom', pt_file, source='local')
 
         if self.yolov5 is None:
             raise RuntimeError('Failed to load YOLOv5 model')
@@ -222,12 +220,15 @@ def _post_process(self, pred, origin_h, origin_w, conf_thres, nms_thres):
         # Do nms
         boxes = self._non_max_suppression(
             pred, origin_h, origin_w, conf_thres, nms_thres)
-        result_boxes = boxes[:, :4] if len(boxes) else np.array([])
-        result_scores = boxes[:, 4] if len(boxes) else np.array([])
-        if num_classes > 1 and len(boxes.shape) >= 2:
+        if len(boxes) == 0:
+            return np.empty((0, 4)), np.empty((0,)), np.empty((0,), dtype=int)
+
+        result_boxes = boxes[:, :4]
+        result_scores = boxes[:, 4]
+        if num_classes > 1:
             result_classid = np.argmax(boxes[:, 5:], axis=1)
         else:
-            # Either we have no classes or all boxes have been removed in _non_max_suppression
+            # Without classes there is no classid to return
             result_classid = np.zeros(boxes.shape[0], dtype=int)
         return result_boxes, result_scores, result_classid
 
@@ -251,7 +252,7 @@ def _non_max_suppression(self, pred, origin_h, origin_w, conf_thres, nms_thres):
         if len(boxes) == 0:
             return np.array([])
         num_classes = boxes.shape[1] - 5
-        # Trasform bbox from [center_x, center_y, w, h] to [x1, y1, x2, y2]
+        # Transform bbox from [center_x, center_y, w, h] to [x1, y1, x2, y2]
         boxes[:, :4] = self.xywh2xyxy(origin_h, origin_w, boxes[:, :4])
         # Object confidence
         confs = boxes[:, 4]