rknn c++ api与python api运算结果不一致，但如果输入参数为0一致

[复制链接] · 发表于 2024-5-15 12:48:14

rt，我在安卓平板上部署rknn算法，通过python运行的推理结果和通过android部署的c++代码运行结果不一致，但如果输入参数为0则一致，求解
截图：
全部参数为0：

python 代码

import os

import urllib

import traceback

import time

import sys

import numpy as np

import cv2

from rknn.api import RKNN

from blazefacenumpy import BlazeFace

ONNX_MODEL = 'best.onnx' # onnx 模型的路径

ONNX_MODEL = 'rppg_new.onnx' # onnx 模型的路径

ONNX_MODEL = 'blazeface128.onnx' # onnx 模型的路径

# ONNX_MODEL = 'v3.onnx' # onnx 模型的路径

# ONNX_MODEL = 'alexnet.onnx' # onnx 模型的路径

TFLITE_MODEL = 'face_detection_front.tflite'

# TFLITE_MODEL = 'alexnet_float32.tflite'





RKNN_MODEL = './yolov8-ghost-pose.rknn'  # 转换后的 RKNN 模型保存路径

RKNN_MODEL = './bz128_3568.rknn'  # 转换后的 RKNN 模型保存路径

# RKNN_MODEL = './facenet——3568.rknn'  # 转换后的 RKNN 模型保存路径

# RKNN_MODEL = './pfldv3.rknn'  # 转换后的 RKNN 模型保存路径

DATASET = './test.txt'   # 数据集文件路径

 

QUANTIZE_ON = False   # 是否进行量化



if __name__ == '__main__':

 

        # 创建 RKNN 对象

        rknn = RKNN(verbose=False)

 

    # 检查 ONNX 模型文件是否存在

        if not os.path.exists(ONNX_MODEL):

                print('model not exist')

                exit(-1)

 

        # 配置模型预处理参数

        print('--> Config model')

        rknn.config(#reorder_channel='0 1 2', # 表示 RGB 通道

                                

                            mean_values=[[128, 128, 128]], # 每个通道的像素均值，预处理时对应通道减去该值

                            std_values=[[128, 128, 128]], # 每个通道的像素标准差，每个通道除以该值

                            optimization_level=3, # 优化级别

                                quantized_method='layer',

                                float_dtype="float16",

                            target_platform = 'RK3568', #指定目标平台为rv1126

                                # quantize_input_node=QUANTIZE_ON

                                )  # 对时输入节点进行量化

        print('done')

 

        # 加载 ONNX 模型

        # print('--> Loading model')

        model = rknn.load_rknn(RKNN_MODEL)

        ret = rknn.init_runtime(target="rk3568")



        frame = cv2.imread('./test.jpg')

        frame =cv2.resize(frame,(128,128))





        npp = np.full_like(frame,2)



        # color_img = np.zeros([1,3,128,128]).astype(np.float16)

        pred = rknn.inference(inputs=[npp], data_format='nhwc',inputs_pass_through=[False])

        # ret = rknn.load_onnx(model=ONNX_MODEL)

        # rknn.inference(inputs=[])

        exit(1)
复制代码

c++ 代码

#include <android/log.h>

#include <android/bitmap.h>



#include "../rknn_api.h"

#include "im2d_type.h"

#include <sys/time.h>

#include <string>

#include <vector>

#include <jni.h>





#include "../rga/rga.h"

#include "../rga/im2d.h"

#include "../rga/im2d_version.h"



#define LOGE(...) __android_log_print(ANDROID_LOG_ERROR, "wuzhengsdkC++", ##__VA_ARGS__);

#define LOGI(...) __android_log_print(ANDROID_LOG_INFO, "wuzhengsdkC++", ##__VA_ARGS__);





uint32_t n_input = 1;

uint32_t n_output = 2;



int m_in_width = 0;   // the width of the RKNN model input

int m_in_height = 0;  // the height of the RKNN model input

int m_in_channel = 0; // the channel of the RKNN model input



rknn_tensor_attr input_attrs[1];

rknn_tensor_attr output_attrs[2];



rknn_tensor_mem *output_mems[2];



std::vector<float> out_scales;

std::vector<int32_t> out_zps;



bool created = false;

rknn_context ctx = 0;



//rga_buffer_t g_rga_dst;

int g_inf_count = 0;



char *jstringToChar(JNIEnv *env, jstring jstr) {

    char *rtn = NULL;

    jclass clsstring = env->FindClass("java/lang/String");

    jstring strencode = env->NewStringUTF("GB2312");

    jmethodID mid = env->GetMethodID(clsstring, "getBytes", "(Ljava/lang/String;)[B");

    jbyteArray barr = (jbyteArray) env->CallObjectMethod(jstr, mid, strencode);

    jsize alen = env->GetArrayLength(barr);

    jbyte *ba = env->GetByteArrayElements(barr, JNI_FALSE);

    if (alen > 0) {

        rtn = (char *) malloc(alen + 1);

        memcpy(rtn, ba, alen);

        rtn[alen] = 0;

    }

    env->ReleaseByteArrayElements(barr, ba, 0);

    return rtn;

}





JNIEXPORT void JNI_OnUnload(JavaVM *vm, void *reserved) {

    LOGE("JNI_OnUnload");

}



extern "C"

JNIEXPORT jint JNICALL Java_com_wuzhengai_examination_tfliterun_blazeface_rknn_BlazeFaceRknn_init(

        JNIEnv *env,

        jobject thiz,

        jstring jmodel_path) {





    char *model_path = jstringToChar(env, jmodel_path);





    rknn_sdk_version version;

    int ret1 = rknn_query(ctx, RKNN_QUERY_SDK_VERSION, &version,

                          sizeof(rknn_sdk_version));

//    printf("api version   : %s\n", version.api_version);

//    printf("driver version: %s\n", version.drv_version);

    LOGE("api version   : %s\n", (char *) version.api_version);

    LOGE("driver version: %s\n", (char *) version.drv_version);

    LOGE("%s,%s", version.api_version, version.drv_version)



    // 1. Load model

    FILE *fp = fopen(model_path, "rb");

    if (fp == NULL) {

        LOGE("fopen %s fail!\n", model_path);

        return -1;

    }

    delete model_path;

    model_path = NULL;

    fseek(fp, 0, SEEK_END);

    uint32_t model_len = ftell(fp);

    void *model = malloc(model_len);

    fseek(fp, 0, SEEK_SET);

    if (model_len != fread(model, 1, model_len, fp)) {

        LOGE("fread %s fail!\n", model_path);

        free(model);

        fclose(fp);

        return -1;

    }

    fclose(fp);



    // 2. Init RKNN model

    int ret = rknn_init(&ctx, model, model_len, 0, nullptr);

    free(model);

    if (ret < 0) {

        LOGE("rknn_init fail! ret=%d\n", ret);

        return -1;

    }



    // 3. Query input/output attr.

    rknn_input_output_num io_num;

    rknn_query_cmd cmd = RKNN_QUERY_IN_OUT_NUM;

    // 3.1 Query input/output num.

    ret = rknn_query(ctx, cmd, &io_num, sizeof(io_num));

    if (ret != RKNN_SUCC) {

        LOGE("rknn_query io_num fail!ret=%d\n", ret);

        return -1;

    }

    n_input = io_num.n_input;

    n_output = io_num.n_output;





    // 3.2 Query input attributes

    memset(input_attrs, 0, n_input * sizeof(rknn_tensor_attr));

    for (int i = 0; i < n_input; ++i) {

        input_attrs[i].index = i;

        cmd = RKNN_QUERY_INPUT_ATTR;

        ret = rknn_query(ctx, cmd, &(input_attrs[i]), sizeof(rknn_tensor_attr));



        if (ret < 0) {

            LOGE("rknn_query input_attrs[%d] fail!ret=%d\n", i, ret);

            return -1;

        }

    }



    // 3.2.0 Update global model input shape.

    if (RKNN_TENSOR_NHWC == input_attrs[0].fmt) {

        m_in_height = input_attrs[0].dims[1];

        m_in_width = input_attrs[0].dims[2];

        m_in_channel = input_attrs[0].dims[3];

    } else if (RKNN_TENSOR_NCHW == input_attrs[0].fmt) {

        m_in_height = input_attrs[0].dims[2];

        m_in_width = input_attrs[0].dims[3];

        m_in_channel = input_attrs[0].dims[1];

    } else {

        LOGE("Unsupported model input layout: %d!\n", input_attrs[0].fmt);

        return -1;

    }







    // 3.3 Query output attributes

    memset(output_attrs, 0, n_output * sizeof(rknn_tensor_attr));

    for (int i = 0; i < n_output; ++i) {

        output_attrs[i].index = i;

        cmd = RKNN_QUERY_OUTPUT_ATTR;

        ret = rknn_query(ctx, cmd, &(output_attrs[i]), sizeof(rknn_tensor_attr));

        if (ret < 0) {

            LOGE("rknn_query output_attrs[%d] fail!ret=%d\n", i, ret);

            return -1;

        }

        // set out_scales/out_zps for post_process

        out_scales.push_back(output_attrs[i].scale);

        out_zps.push_back(output_attrs[i].zp);

    }



    void *in_data = malloc(m_in_width * m_in_height * m_in_channel);

    memset(in_data, 0, m_in_width * m_in_height * m_in_channel);



    // 4.1.2 Update input attrs

    input_attrs[0].index = 0;

    input_attrs[0].type = RKNN_TENSOR_UINT8;

    input_attrs[0].size = m_in_height * m_in_width * m_in_channel * sizeof(char);

    input_attrs[0].fmt = RKNN_TENSOR_NCHW;

    // TODO -- The efficiency of pass through will be higher, we need adjust the layout of input to

    //         meet the use condition of pass through.

    input_attrs[0].pass_through = 0;

    // 4.1.3 Set input buffer

//    rknn_set_io_mem(ctx, input_mems[0], &(input_attrs[0]));



    created = true;



    return 0;

}







extern "C"

JNIEXPORT void JNICALL

Java_com_wuzhengai_examination_tfliterun_blazeface_rknn_BlazeFaceRknn_release(

        JNIEnv *env,

        jobject thiz) {

    LOGI("rknn_destroy!");

    // release io_mem resource

//    for (int i = 0; i < n_input; ++i) {

//        rknn_destroy_mem(ctx, input_mems[i]);

//    }

    for (int i = 0; i < n_output; ++i) {

        rknn_destroy_mem(ctx, output_mems[i]);

    }

    rknn_destroy(ctx);



}







extern "C"

JNIEXPORT jboolean JNICALL

Java_com_wuzhengai_examination_tfliterun_blazeface_rknn_BlazeFaceRknn_detect(

        JNIEnv *env,

        jobject thiz,

        jintArray input,

        jfloatArray output1,

        jfloatArray output2) {



    if (!created) {

        LOGE("run_yolo: init yolo hasn't successful!");

        return false;

    }



    jboolean inputCopy = JNI_FALSE;

    jint *const inData = env->GetIntArrayElements(input, &inputCopy);



    // 获取数组长度

    jsize length = env->GetArrayLength(input);



// 创建一个int8_t数组

    uint8_t *inputData = new uint8_t[length];

    // inputData

    for (int i = 0; i < length; i++) {

        inputData[i] = static_cast<uint8_t>(inData[i]);

    }





    LOGE("test,%d,%d,%d", inData[126],static_cast<uint8_t>(inData[126]),inputData[126])

    jboolean outputCopy = JNI_FALSE;





    jfloat *const y0 = env->GetFloatArrayElements(output1, &outputCopy);

    jfloat *const y1 = env->GetFloatArrayElements(output2, &outputCopy);





    rknn_input inputs[1];

    memset(inputs, 0, sizeof(inputs));

    inputs[0].index = 0;

    inputs[0].type = RKNN_TENSOR_UINT8;

//    float 类型需要乘以2?不清楚

    inputs[0].size = m_in_height * m_in_width * m_in_channel ;

//    inputs[0].size =75264;

    inputs[0].fmt = RKNN_TENSOR_NHWC;

    inputs[0].pass_through = 0;

    inputs[0].buf = inData;





    int ret = rknn_inputs_set(ctx, 1, inputs);





    rknn_output outputs[2];

    memset(outputs, 0, sizeof(outputs));

    for (int i = 0; i < 2; ++i) {

        outputs[i].want_float = 1;

    }





    if (ret < 0) {

        LOGE("rknn_set fail! ret=%d\n", ret);

        return false;

    }

//    LOGE("rknn set %d",ret)

    ret = rknn_run(ctx, nullptr);

    if (ret < 0) {

        LOGE("rknn_run fail! ret=%d\n", ret);

        return false;

    }





    rknn_outputs_get(ctx, 2, outputs, NULL);



    memcpy(y0, outputs[0].buf, output_attrs[0].n_elems * sizeof(char));

    memcpy(y1, outputs[1].buf, output_attrs[1].n_elems * sizeof(char));



//

    env->ReleaseIntArrayElements(input, inData, JNI_ABORT);

    env->ReleaseFloatArrayElements(output1, y0, 0);

    env->ReleaseFloatArrayElements(output2, y1, 0);



    rknn_outputs_release(ctx, 2, outputs);

    return JNI_TRUE;

}



复制代码

只看该作者 · 发表于 2024-5-15 12:54:58

这是0输入，细看好像也不是完全一样。。

奇怪？怎么图片好像上不来？