meta-rknn 0.1.2

Meta RKNN Toolkit

metarknn

metarknn部署通用框架

1、安装最新版 meta-cv

pip install meta-cv

2、安装最新版 meta-rknn

pip install meta-rknn

3、图像分类示例（参考classification_demo.py代码）

import cv2, platform
import metarknn as m

Classification = m.Classification

y = Classification(model_path='models/mixnet_xl_bs24_2.rknn',
                   input_width=224,
                   input_height=224,
                   use_preprocess=True,
                   class_names=classnames,
                   device_id=0)

batch_size = 24
img = cv2.imread('models/bottle.jpg')
img_list = [img[:, :, ::-1]] * batch_size if batch_size > 1 else img[:, :, ::-1]
_dets, _scores, _labels = y.predict(img_list)

# 显示
y.show(img, _dets, _scores, _labels)
cv2.imwrite("models/bottle.png", img)

4、目标检测示例（参考detection_demo.py代码）

import cv2, platform
import metarknn as m

Detection = m.Detection

y = Detection(model_path='models/yolov8m.rknn',
              input_width=640,
              input_height=480,
              use_preprocess=True,
              pad=True,
              confidence_thresh=0.5,
              nms_thresh=0.3,
              class_names=classnames,
              device_id=0)

batch_size = 1
img = cv2.imread('models/bus.jpg')
img_list = [img[:, :, ::-1]] * batch_size if batch_size > 1 else img[:, :, ::-1]
_dets, _scores, _labels = y.predict(img_list)

# 显示
y.show(img, _dets[-1], _scores[-1], _labels[-1])
cv2.imwrite("models/bus.png", img)

5、实例分割示例（参考segment_demo.py代码）

import platform, cv2
import metarknn as m

Segment = m.Segment

y = Segment(model_path='models/yolov8m-seg.rknn',
            input_width=640,
            input_height=480,
            use_preprocess=True,
            pad=True,
            confidence_thresh=0.5,
            nms_thresh=0.3,
            class_names=classnames,
            device_id=0)

batch_size = 1
img = cv2.imread('models/bus.jpg')
img_list = [img[:, :, ::-1]] * batch_size if batch_size > 1 else img[:, :, ::-1]
_dets, _scores, _labels = y.predict(img_list)

# 显示
y.show(img, _dets[-1], _scores[-1], _labels[-1])
cv2.imwrite("models/bus.png", img)

6、模型转换与量化（本地运行）

import metarknn as m

Quantization = m.Quantization

q = Quantization(model_path，    # onnx模型路径
                 dataset,   # dataset文件路径
                 mean=[[0, 0, 0]],
                 std=[[255, 255, 255]],
                 batch_size=1)   # 定义模型输出层

if is_hybrid:   # 是否混合量化
    self.model = q.hybrid_convert()
else:   # 非混合量化（是否int8量化）
    self.model = q.convert(quantize)