[Feature] Support YOLOv5 deployment at RV1126 device (open-mmlab#262)

* add yolov5 head rewrite and rknn config

* type hint

* remove useless

* refactor config

configs/deploy/detection_rknn-fp16_static-320x320.py

@@ -0,0 +1,9 @@
+_base_ = ['./base_static.py']
+onnx_config = dict(
+    input_shape=[320, 320], output_names=['feat0', 'feat1', 'feat2'])
+codebase_config = dict(model_type='rknn')
+backend_config = dict(
+    type='rknn',
+    common_config=dict(target_platform='rv1126', optimization_level=1),
+    quantization_config=dict(do_quantization=False, dataset=None),
+    input_size_list=[[3, 320, 320]])

configs/deploy/detection_rknn-int8_static-320x320.py

@@ -0,0 +1,9 @@
+_base_ = ['./base_static.py']
+onnx_config = dict(
+    input_shape=[320, 320], output_names=['feat0', 'feat1', 'feat2'])
+codebase_config = dict(model_type='rknn')
+backend_config = dict(
+    type='rknn',
+    common_config=dict(target_platform='rv1126', optimization_level=1),
+    quantization_config=dict(do_quantization=True, dataset=None),
+    input_size_list=[[3, 320, 320]])

mmyolo/deploy/models/dense_heads/yolov5_head.py

@@ -146,3 +146,34 @@ def yolov5_head__predict_by_feat(ctx,
 
     return nms_func(bboxes, scores, max_output_boxes_per_class, iou_threshold,
                     score_threshold, pre_top_k, keep_top_k)
+
+
+@FUNCTION_REWRITER.register_rewriter(
+    func_name='mmyolo.models.dense_heads.yolov5_head.'
+    'YOLOv5Head.predict',
+    backend='rknn')
+def yolov5_head__predict__rknn(ctx, self, x: Tuple[Tensor], *args,
+                               **kwargs) -> Tuple[Tensor, Tensor, Tensor]:
+    """Perform forward propagation of the detection head and predict detection
+    results on the features of the upstream network.
+
+    Args:
+        x (tuple[Tensor]): Multi-level features from the
+            upstream network, each is a 4D-tensor.
+    """
+    outs = self(x)
+    return outs
+
+
+@FUNCTION_REWRITER.register_rewriter(
+    func_name='mmyolo.models.dense_heads.yolov5_head.'
+    'YOLOv5HeadModule.forward',
+    backend='rknn')
+def yolov5_head_module__forward__rknn(
+        ctx, self, x: Tensor, *args,
+        **kwargs) -> Tuple[Tensor, Tensor, Tensor]:
+    """Forward feature of a single scale level."""
+    out = []
+    for i, feat in enumerate(x):
+        out.append(self.convs_pred[i](feat))
+    return out