使用RK3588的NPU进行Yolo-V5推理
YOU Only Look Once--
BUT I Look for Many Times
使用RK3588NPU进行Yolo-V5推理
前言：
RK3588提供了6Tops的计算性能，能够很好的进行Yolo-v5的推理，本文参考了官方的实例，对代码结构进行简化，增强了可读性，便于初学者的理解。
环境配置：
安装项目依赖：
pip install opencv-pythonpip install PyQt下载转换好的
rknn模型
推理流程
初始化
使用Python工具调用推理的api
1
from rknnlite.api import RKNNLite初始化推理对象
1
rknn = RKNNLite()加载模型
1
2yolo = rknn.load_rknn('./rknn/yolov5m_leaky.rknn')
yolo = rknn.init_runtime(core_mask=RKNNLite.NPU_CORE_0_1_2) # 使用所有的NPU核心
进行推理
使用rknn.inferience(input=[img])接口进行推理
编写推理逻辑
读取待处理文件
1 | |
处理输出信息
推理的结果包含着255个不同维度的信息，分别包含了锚框的坐标信息、置信度分数、识别的物体类别，因此，可以将多维度分解为
3个batch,使得输出boxes, scores, class有足够的信息筛选锚框：
使用对象抑制，筛选掉置信度低的锚框——去除识别错误的物体
1
2
3
4
5
6
7
8
9
10
11
12
13def filter_boxes(boxes, box_confidences, box_class_probs, ObjectThreshold=0.25):
box_confidences = box_confidences.reshape(-1)
class_max_score = np.max(box_class_probs, axis=-1)
classes = np.argmax(box_class_probs, axis=-1)
confidence = np.multiply(class_max_score,box_confidences)
_class_pos = np.where(confidence >= ObjectThreshold)
scores = confidence[_class_pos]
boxes = boxes[_class_pos]
classes = classes[_class_pos]
return boxes, classes, scores使用NMS抑制，筛选出识别同一个物体效果最好的锚块
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28def nms_boxes(boxes, scores, NMS_THRESH=0.45):
x = boxes[:, 0]
y = boxes[:, 1]
w = boxes[:, 2] - boxes[:, 0]
h = boxes[:, 3] - boxes[:, 1]
areas = w * h
order = scores.argsort()[::-1]
keep = []
while order.size > 0:
i = order[0]
keep.append(i)
xx1 = np.maximum(x[i], x[order[1:]])
yy1 = np.maximum(y[i], y[order[1:]])
xx2 = np.minimum(x[i] + w[i], x[order[1:]] + w[order[1:]])
yy2 = np.minimum(y[i] + h[i], y[order[1:]] + h[order[1:]])
w1 = np.maximum(0.0, xx2 - xx1 + 0.00001)
h1 = np.maximum(0.0, yy2 - yy1 + 0.00001)
inter = w1 * h1
ovr = inter / (areas[i] + areas[order[1:]] - inter)
inds = np.where(ovr <= NMS_THRESH)[0]
order = order[inds + 1]
keep = np.array(keep)
return keep
连接模块，将推理结果转化为信息，返回
boxes shape=(n x 4)，classes shape=(n x 1)，scores shape=(n x 1)流程：
1. 输入255维度，转化为`3*85` 2. 前4个维度预测锚框的坐标，第5个维度预测置信度分数，后面的维度用于进行类别预测1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58import numpy as np
from . import Tools
from . import generateBox
default_anchors = np.array([ 10., 13., 16., 30., 33., 23., 30., 61., 62., 45., 59., 119., 116., 90., 156., 198., 373., 326.])
default_anchors = default_anchors.reshape(3,-1,2).tolist()
def post_process(input_data, anchors=default_anchors):
boxes, scores, classes_conf = [], [], []
# 1*255*h*w -> 3*85*h*w
input_data = [_in.reshape([
len(anchors[0]),-1]+
list(_in.shape[-2:]))
for _in in input_data]
for i in range(len(input_data)):
boxes.append(generateBox.box_process(input_data[i][:,:4,:,:], anchors[i]))
scores.append(input_data[i][:,4:5,:,:])
classes_conf.append(input_data[i][:,5:,:,:])
def sp_flatten(_in):
ch = _in.shape[1]
_in = _in.transpose(0,2,3,1)
return _in.reshape(-1, ch)
boxes = [sp_flatten(_v) for _v in boxes]
classes_conf = [sp_flatten(_v) for _v in classes_conf]
scores = [sp_flatten(_v) for _v in scores]
boxes = np.concatenate(boxes)
classes_conf = np.concatenate(classes_conf)
scores = np.concatenate(scores)
# filter according to threshold
boxes, classes, scores = Tools.filter_boxes(boxes, scores, classes_conf)
# nms
nboxes, nclasses, nscores = [], [], []
for c in set(classes):
inds = np.where(classes == c)
b = boxes[inds]
c = classes[inds]
s = scores[inds]
keep = Tools.nms_boxes(b, s)
if len(keep) != 0:
nboxes.append(b[keep])
nclasses.append(c[keep])
nscores.append(s[keep])
if not nclasses and not nscores:
return None, None, None
boxes = np.concatenate(nboxes)
classes = np.concatenate(nclasses)
scores = np.concatenate(nscores)
return boxes, classes, scores可视化展示
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40import time
import cv2
import numpy as np
from . import OutputPredict
CLASSES = ("person", "bicycle", "car","motorbike ","aeroplane ","bus ","train","truck ","boat","traffic light",
"fire hydrant","stop sign ","parking meter","bench","bird","cat","dog ","horse ","sheep","cow","elephant",
"bear","zebra ","giraffe","backpack","umbrella","handbag","tie","suitcase","frisbee","skis","snowboard","sports ball","kite",
"baseball bat","baseball glove","skateboard","surfboard","tennis racket","bottle","wine glass","cup","fork","knife ",
"spoon","bowl","banana","apple","sandwich","orange","broccoli","carrot","hot dog","pizza ","donut","cake","chair","sofa",
"pottedplant","bed","diningtable","toilet ","tvmonitor","laptop ","mouse ","remote ","keyboard ","cell phone","microwave ",
"oven ","toaster","sink","refrigerator ","book","clock","vase","scissors ","teddy bear ","hair drier", "toothbrush ")
def video_predict(path,rknn):
cap = cv2.VideoCapture(path)
while cap.isOpened():
start_time = time.time()
ret, frame = cap.read()
if not ret:
break
output = rknn.inference(inputs=[np.expand_dims(frame, axis=0)])
boxes, classes, scores = OutputPredict.post_process(output)
end_time = time.time()
fps = 1 / (end_time - start_time)
if boxes is not None and classes is not None and scores is not None:
classes_out = [CLASSES[i] for i in classes]
scores_out = [f"{_ * 100:.1f}%" for _ in scores]
for box, text_cl, text_sc in zip(boxes, classes_out, scores_out):
box = box.astype(int)
x1, y1, x2, y2 = box
text = f"{text_cl}, {text_sc}, FPS:{fps:.1f}"
frame = cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
cv2.putText(frame, text, (x1, y1-10),
fontFace=cv2.FONT_ITALIC,
fontScale=0.5, color=(0, 0, 255),
thickness=1)
cv2.imshow('Output', frame)
if cv2.waitKey(1) == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
项目演示
据说
RKNN的Python接口性能优化不好，NPU性能无法完全调用，双路推理不足20帧，还有很大的优化空间，目前完成度仅限使用了NPU查看NPU占用
1
watch -0.5 cat /sys/kernel/debug/rknpu/load