Mediapipe 手势模型 转换rknn

mediappipe手势模型低端设备  RK3568推理 GPU 占用比较高  尝试3568平台NPU 进行推理 

一下进行 tflite模型转换 rknn模型

手部坐标 模型  图

准备rknn-tookit2 1.3.0以上环境  

进入 t@ubuntu:~/rknn/rknn-toolkit2/examples/tflite$ 目录

参照 mobilenet_v1 demo 转换hands模型

copy mobilenet_v1 demo  重命名 mediapipe_hand

cd /rknn/rknn-toolkit2/examples/tflite/mediapipe_hand$

修改 test.py  加载模型 及输入图片资源 

import numpy as np
import cv2
from rknn.api import RKNN
 
#import tensorflow.compat.v1 as tf #使用1.0版本的方法
#tf.disable_v2_behavior() #
 
def show_outputs(outputs):
    output = outputs[0][0]
    output_sorted = sorted(output, reverse=True)
    top5_str = 'mobilenet_v1\n-----TOP 5-----\n'
    for i in range(5):
        value = output_sorted[i]
        index = np.where(output == value)
        for j in range(len(index)):
            if (i + j) >= 5:
                break
            if value > 0:
                topi = '{}: {}\n'.format(index[j], value)
            else:
                topi = '-1: 0.0\n'
            top5_str += topi
    print(top5_str)
 
 
if __name__ == '__main__':
 
    # Create RKNN object
    rknn = RKNN(verbose=True)
 
    # Pre-process config
    print('--> Config model')
    rknn.config(mean_values=[128, 128, 128], std_values=[128, 128, 128])
    print('done')
 
    # Load model
    print('--> Loading model')
    ret = rknn.load_tflite(model='hand_landmark_lite.tflite')
    #ret = rknn.load_tflite(model='palm_detection_lite.tflite')
    if ret != 0:
        print('Load model failed!')
        exit(ret)
    print('done')
 
    # Build model
    print('--> Building model')
    ret = rknn.build(do_quantization=True, dataset='./dataset.txt')
    if ret != 0:
        print('Build model failed!')
        exit(ret)
    print('done')
 
    # Export rknn model
    print('--> Export rknn model')
    ret = rknn.export_rknn('./hands.rknn')
    if ret != 0:
        print('Export rknn model failed!')
        exit(ret)
    print('done')
 
    # Set inputs
    img = cv2.imread('./hand_1.jpg')
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, (224,224))
    img = np.expand_dims(img, 0)
 
    # Init runtime environment
    print('--> Init runtime environment')
    ret = rknn.init_runtime()
    if ret != 0:
        print('Init runtime environment failed!')
        exit(ret)
    print('done')
 
    # Inference
    print('--> Running model')
    outputs = rknn.inference(inputs=[img])
    print('--> outputs:' ,outputs)
    np.save('./tflite_hands.npy', outputs[0])
    show_outputs(outputs)
    print('done')
 
    rknn.release()

注意输入 图的size    tensorflow版本 尽量使用最新的  

模拟推理结果

数据格式

[array([[ 75.521454, 161.8317 , 0. , 110.37751 , 165.15132 ,
-6.639249, 141.91394 , 159.34198 , -11.618686, 170.13075 ,
151.04291 , -15.768216, 190.0485 , 141.91394 , -19.917747,
131.95508 , 99.58873 , -14.108404, 158.51207 , 63.90277 ,
-18.257935, 174.28029 , 45.644836, -19.917747, 186.72887 ,
31.536432, -20.747652, 112.86723 , 87.97005 , -12.448591,
136.1046 , 48.134556, -15.768216, 151.04291 , 24.067278,
-17.428028, 164.32141 , 7.469155, -17.428028, 96.26911 ,
85.48033 , -9.128967, 109.54761 , 47.30465 , -12.448591,
120.33639 , 25.727089, -13.278498, 130.29526 , 10.788779,
-13.278498, 81.330795, 87.97005 , -7.469155, 83.82052 ,
55.60371 , -9.128967, 90.45976 , 38.175682, -9.958874,
97.92892 , 24.897182, -9.958874]], dtype=float32), array([[0.9889264]], dtype=float32), array([[0.4299424]], dtype=float32), array([[-0.03374431, 0.06308718, 0.01027001, 0.00293429, 0.0572186 ,
0.02689764, 0.02200716, 0.04645955, 0.01369334, 0.04059098,
0.04743765, 0.01075905, 0.06210908, 0.03863478, -0.00978096,
0.01907287, -0.0009781 , -0.0009781 , 0.03521145, -0.01075905,
-0.01075905, 0.04010193, -0.0224962 , 0.00489048, 0.05672956,
-0.03618954, 0.02738668, 0.00342334, -0.00244524, -0.00391238,
0.01516049, -0.02689764, -0.00146714, 0.02640859, -0.05183908,
0.02151811, 0.04059098, -0.06113099, 0.00537953, -0.00391238,
-0.00244524, 0.0009781 , -0.00635762, -0.02200716, -0.00146714,
0.00244524, -0.04401431, 0.00489048, 0.01369334, -0.0581967 ,
0.02396335, -0.01467144, -0.00880286, 0.00586857, -0.02836478,
-0.01956192, 0.00244524, -0.02151811, -0.03716764, -0.00048905,
-0.01173715, -0.04499241, -0.01907287]], dtype=float32)]

 手部坐标预测模型 （Hand LandMark Model）

21个关键点 hands检测置信度 左右手分类 ..