完善资料让更多小伙伴认识你，还能领取20积分哦， 立即完善>
3天内不再提示
3584
|
1.模型转换
1.1 yolov5-face的pt模型转为onnx模型 从github上下载yolov5-face工程，然后利用里面的exprot.py将pytorch模型转换为onnx模型，转换之前需要修改yolo.py代码，从github上下载yolov5-face工程，然后利用里面的exprot.py将pytorch模型转换为onnx模型，转换之前需要修改yolo.py代码， 1. def forward(self, x): 2. # x = x.copy() # for profiling 3. z = [] # inference output 4. # self.training |= self.export 5. if self.export: 6. for i in range(self.nl): 7. x = self.m(x) 8. bs, _, ny, nx = x.shape # x(bs,48,20,20) to x(bs,3,20,20,16) 9. x = x.view(bs, self.na, self.no, ny, nx). permute(0, 1, 3, 4, 2).contiguous() 需要将代码中的 .permute(0, 1, 3, 4, 2)去掉，修改为如下代码 1. def forward(self, x): 2. # x = x.copy() # for profiling 3. z = [] # inference output 4. # self.training |= self.export 5. if self.export: 6. for i in range(self.nl): 7. x = self.m(x) 8. bs, _, ny, nx = x.shape # x(bs,48,20,20) to x(bs,3,20,20,16) 9. x = x.view(bs, self.na, self.no, ny, nx).contiguous() 这样修改的原因是后面我们的C++代码解析的是3*16*20*20类型的，而不是3*20*20*16类型的。 1.2 yolov5-face的onnx模型转换为rknn模型 由于rknntoolkit1.6.0在转换yolov-face的模型时报错，因此首先需要安装toolkit1.7.1，然后再用https://github.com/airockchip/yolov5/tree/master/rknn中的onnx2rknn.py进行模型转换，转换脚本如下。 1. import os 2. import sys 3. import numpy as np 4. from rknn.api import RKNN 5. 6. ONNX_MODEL = 'yolov5s-face.onnx' 7. RKNN_MODEL = 'yolov5s-face.rknn' 8. 9. if __name__ == '__main__': 10. 11. # Create RKNN object 12. rknn = RKNN(verbose=True) 13. 14. # pre-process config 15. print('--> config model') 16. rknn.config(mean_values=[[0, 0, 0]], std_values=[[255, 255, 255]], reorder_channel='0 1 2', target_platform='rv1126', 17. quantized_dtype='asymmetric_affine-u8', optimization_level=3, output_optimize=1) 18. print('done') 19. 20. print('--> Loading model') 21. ret = rknn.load_onnx(model=ONNX_MODEL) 22. if ret != 0: 23. print('Load model failed!') 24. exit(ret) 25. print('done') 26. 27. # Build model 28. print('--> Building model') 29. ret = rknn.build(do_quantization=True, dataset='./dataset.txt') 30. if ret != 0: 31. print('Build yolov5s failed!') 32. exit(ret) 33. print('done') 34. 35. # Export rknn model 36. print('--> Export RKNN model') 37. ret = rknn.export_rknn(RKNN_MODEL) 38. if ret != 0: 39. print('Export yolov5s.rknn failed!') 40. exit(ret) 41. print('done') 42. 43. rknn.release() 2.C++代码 这里以瑞芯微的代码为例，只不过在前面和后面加上了有关padding resize的处理，瑞芯微的具体代码见rknpu/rknn/rknn_api/examples/rknn_yolov5_demo at master · rockchip-linux/rknpu · GitHub 2.1 padding resize代码 瑞芯微的官方demo里面是用opencl读取图片，这里我改成了用opencv读取图片，关于opencv的交叉编译见：ubuntu交叉编译RV1126的opencv库/ubuntu交叉编译opencv void padding_resize(cv::InputArray src, cv::OutputArray dst, cv::Size size) { float padd_w = 0; float padd_h = 0; float r = std::min(float(size.width) / src.cols(), float(size.height) / src.rows()); int inside_w = round(src.cols() * r); int inside_h = round(src.rows() * r); padd_w = size.width - inside_w; padd_h = size.height - inside_h; cv::resize(src, dst, cv::Size(inside_w, inside_h)); padd_w = padd_w / 2; padd_h = padd_h / 2; //外层边框填充灰色 printf("gain:%f, padd_w:%f,padd_h:%f. in padding_resize============n", r, padd_w,padd_h); int top = int(round(padd_h - 0.1)); int bottom = int(round(padd_h + 0.1)); int left = int(round(padd_w - 0.1)); int right = int(round(padd_w + 0.1)); cv::copyMakeBorder(dst, dst, top, bottom, left, right, cv::BORDER_CONSTANT, cv::Scalar(114, 114, 114)); } cv::Mat read_image(const char *image_path, int w, int h, int & img_width, int & img_height, cv::Mat & img) { img = cv::imread(image_path); img_width = img.cols; img_height = img.rows; cv::Mat sample_resize, sample; padding_resize(img, sample_resize, cv::Size(w, h)); //cv::resize(img, sample_resize, cv::Size(w, h)); cv::cvtColor(sample_resize, sample, cv::COLOR_BGR2RGB); return sample; } 2.2 yolov5-face后处理(在瑞芯微yolov5后处理代码基础上修改) 这是瑞芯微的yolov5 demo里的代码，这里面我把坐标框后处理时影射回原图的四行代码注释掉了。然后其他代码也针对yolov5-face进行了修改，其中process这里增加了关键点的解析。 static int process(uint8_t *input, int *anchor, int grid_h, int grid_w, int height, int width, int stride, std::vector float threshold, uint32_t zp, float scale) { int validCount = 0; int grid_len = grid_h * grid_w;//80*80, 40*40, 20*20. float thres = unsigmoid(threshold); uint8_t thres_u8 = qnt_f32_to_affine(thres, zp, scale);//量化 for (int a = 0; a < 3; a++) { for (int i = 0; i < grid_h; i++) { for (int j = 0; j < grid_w; j++) { /********************************************************************************************** 排列顺序是x y w h boxScore x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 idScore，然后保存的时候是先存 grid_h * grid_w个x，再存grid_h * grid_w个y,....,这样保存的。 而不是先存第一个x y w h boxScore x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 idScore， 再存第二个x y w h boxScore x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 idScore， **********************************************************************************************/ uint8_t box_confidence = input[(PROP_BOX_SIZE * a + 4) * grid_len + i * grid_w + j]; if (box_confidence >= thres_u8)//thres_u8经过反sigmoid和反量化操作了。 { int offset = (PROP_BOX_SIZE * a) * grid_len + i * grid_w + j;// uint8_t *in_ptr = input + offset; /*这边乘以2，减掉0.5是因为内部运算的时候做了这个处理，所以这里要反操作一下, deqnt_affine_to_f32是反量化，zp是量化时的zero point中心点，scale是量化时的尺度。*/ float box_x = sigmoid(deqnt_affine_to_f32(*in_ptr, zp, scale)) * 2.0 - 0.5; float box_y = sigmoid(deqnt_affine_to_f32(in_ptr[grid_len], zp, scale)) * 2.0 - 0.5; float box_w = sigmoid(deqnt_affine_to_f32(in_ptr[2 * grid_len], zp, scale)) * 2.0; float box_h = sigmoid(deqnt_affine_to_f32(in_ptr[3 * grid_len], zp, scale)) * 2.0; float x1 = (deqnt_affine_to_f32(in_ptr[5 * grid_len], zp, scale)); float y1 = (deqnt_affine_to_f32(in_ptr[6 * grid_len], zp, scale)); float x2 = (deqnt_affine_to_f32(in_ptr[7 * grid_len], zp, scale)); float y2 = (deqnt_affine_to_f32(in_ptr[8 * grid_len], zp, scale)); float x3 = (deqnt_affine_to_f32(in_ptr[9 * grid_len], zp, scale)); float y3 = (deqnt_affine_to_f32(in_ptr[10 * grid_len], zp, scale)); float x4 = (deqnt_affine_to_f32(in_ptr[11 * grid_len], zp, scale)); float y4 = (deqnt_affine_to_f32(in_ptr[12 * grid_len], zp, scale)); float x5 = (deqnt_affine_to_f32(in_ptr[13 * grid_len], zp, scale)); float y5 = (deqnt_affine_to_f32(in_ptr[14 * grid_len], zp, scale)); std::cout<<"landmark after deqnt_affine_to_f32:"< box_x = (box_x + j) * (float)stride;//这边的box_x是0-1的，相当于是偏移值，所以要加上j。乘以stride是映射回原图640. box_y = (box_y + i) * (float)stride;//这边的box_y是0-1的，相当于是偏移值，所以要加上i。乘以stride是映射回原图640. box_w = box_w * box_w * (float)anchor[a * 2]; box_h = box_h * box_h * (float)anchor[a * 2 + 1]; box_x -= (box_w / 2.0); box_y -= (box_h / 2.0); boxes.push_back(box_x); boxes.push_back(box_y); boxes.push_back(box_w); boxes.push_back(box_h); x1 = x1 * (float)anchor[a * 2] + j*(float)stride; y1 = y1 * (float)anchor[a * 2 + 1] + i*(float)stride; x2 = x2 * (float)anchor[a * 2] + j*(float)stride; y2 = y2 * (float)anchor[a * 2 + 1] + i*(float)stride; x3 = x3 * (float)anchor[a * 2] + j*(float)stride; y3 = y3 * (float)anchor[a * 2 + 1] + i*(float)stride; x4 = x4 * (float)anchor[a * 2] + j*(float)stride; y4 = y4 * (float)anchor[a * 2 + 1] + i*(float)stride; x5 = x5 * (float)anchor[a * 2] + j*(float)stride; y5 = y5 * (float)anchor[a * 2 + 1] + i*(float)stride; landMarks.push_back(x1); landMarks.push_back(y1); landMarks.push_back(x2); landMarks.push_back(y2); landMarks.push_back(x3); landMarks.push_back(y3); landMarks.push_back(x4); landMarks.push_back(y4); landMarks.push_back(x5); landMarks.push_back(y5); // printf("box_x=%.03f, box_y=%.03f, box_w=%.03f, box_h=%.03fn" // , box_x, box_y, box_w, box_h); uint8_t maxClassProbs = in_ptr[15 * grid_len]; int maxClassId = 0; for (int k = 0; k < OBJ_CLASS_NUM; ++k) { uint8_t prob = in_ptr[(15 + k) * grid_len];//这里是通过比较找到一个最高的得分和他的id。 if (prob > maxClassProbs) { maxClassId = k; maxClassProbs = prob; } } objProbs.push_back(sigmoid(deqnt_affine_to_f32(maxClassProbs, zp, scale))); classId.push_back(maxClassId); //boxes objProbs classId这三个变量往外返回的时候应该是每四个box对应一个objProb和一个classId。 //std::cout<<"maxClassProbs in post_process:::"< } } } } return validCount; } 3 RV1126驱动升级 由于前面使用rknntoolkit 1.7.1进行模型转换的，所以这里需要把RV1126板子的驱动由1.6.1升级为1.7.0，具体方法如下：rv1126更新驱动版本库_cumtchw-CSDN博客 1. 把https://github.com/rockchip-linux/rknpu/tree/master/drivers/linux-armhf-puma/usr push到板子的对应目录。 2. 把https://github.com/rockchip-linux/rknpu/blob/master/drivers/npu_ko/galcore_puma.ko改名为galcore.ko， push到板子/lib/modules/galcore.ko。可以先在板子上find下原来galcore.ko的位置。 3. 重启板子。 升级完成之后需要到/usr/lib目录下用下面三个命令做一下软链接。 1. [root@RV1126_RV1109:/usr/lib]# ln -snf libOpenVX.so.1.2 libOpenVX.so.1 2. [root@RV1126_RV1109:/usr/lib]# ln -snf libOpenVX.so.1 libOpenVX.so 3. [root@RV1126_RV1109:/usr/lib]# ln -snf libOpenCL.so.1.2 libOpenCL.so.1 4.完整的C++代码 4.1 postprocess.h #ifndef _RKNN_ZERO_COPY_DEMO_POSTPROCESS_H_ #define _RKNN_ZERO_COPY_DEMO_POSTPROCESS_H_ #include #include #define OBJ_NAME_MAX_SIZE 16 #define OBJ_NUMB_MAX_SIZE 64 //#define OBJ_CLASS_NUM 2 //80 #define OBJ_CLASS_NUM 1 //yolov5-face的类别数是1. #define NMS_THRESH 0.3 #define BOX_THRESH 0.87 //0.5 // #define NMS_THRESH 0.5 // #define BOX_THRESH 0.3 //0.5 #define LAND_MARK_SIZE 10//五个关键点，每个关键点有两个坐标x,y。 #define PROP_BOX_SIZE (5 + OBJ_CLASS_NUM + LAND_MARK_SIZE) //16 //#define PROP_BOX_SIZE (5+OBJ_CLASS_NUM) //10 #define REPVGG_CLASS_NUM 27 //信号灯分类数 extern char *repvgg_labels[REPVGG_CLASS_NUM]; typedef struct _BOX_RECT { int left; int right; int top; int bottom; //std::vector } BOX_RECT; typedef struct __detect_result_t { char name[OBJ_NAME_MAX_SIZE]; BOX_RECT box; std::vector float prop; } detect_result_t; typedef struct _detect_result_group_t { int id; int count; detect_result_t results[OBJ_NUMB_MAX_SIZE]; } detect_result_group_t; int post_process(uint8_t *input0, uint8_t *input1, uint8_t *input2, int model_in_h, int model_in_w, float conf_threshold, float nms_threshold, float scale_w, float scale_h, std::vector detect_result_group_t *group); int rknn_GetTop( float *pfProb, float *pfMaxProb, uint32_t *pMaxClass, uint32_t outputCount, uint32_t topNum); int loadLabelName(const char *locationFilename, char *label[], int classNum); #endif //_RKNN_ZERO_COPY_DEMO_POSTPROCESS_H_ 4.2 postprocess.cc // Copyright (c) 2021 by Rockchip Electronics Co., Ltd. All Rights Reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. #include #include #include #include #include #include #include #include "postprocess.h" #include #define LABEL_NALE_TXT_PATH "./model/coco_80_labels_list.txt" static char *labels[OBJ_CLASS_NUM]; char *repvgg_labels[REPVGG_CLASS_NUM]; //这是之前yolov5的anchor // const int anchor0[6] = {10, 13, 16, 30, 33, 23}; // const int anchor1[6] = {30, 61, 62, 45, 59, 119}; // const int anchor2[6] = {116, 90, 156, 198, 373, 326}; //这是yolov5-face的anchor. const int anchor0[6] = {4, 5, 8, 10, 13, 16}; const int anchor1[6] = {23, 29, 43, 55, 73, 105}; const int anchor2[6] = {146, 217, 231, 300, 335, 433}; inline static int clamp(float val, int min, int max) { return val > min ? (val < max ? val : max) : min; } char *readLine(FILE *fp, char *buffer, int *len) { int ch; int i = 0; size_t buff_len = 0; buffer = (char *)malloc(buff_len + 1); if (!buffer) return NULL; // Out of memory while ((ch = fgetc(fp)) != 'n' && ch != EOF) { buff_len++; void *tmp = realloc(buffer, buff_len + 1); if (tmp == NULL) { free(buffer); return NULL; // Out of memory } buffer = (char *)tmp; buffer = (char)ch; i++; } buffer = ' '; *len = buff_len; // Detect end if (ch == EOF && (i == 0 || ferror(fp))) { free(buffer); return NULL; } return buffer; } int readLines(const char *fileName, char *lines[], int max_line) { FILE *file = fopen(fileName, "r"); char *s; int i = 0; int n = 0; while ((s = readLine(file, s, &n)) != NULL) { lines[i++] = s; if (i >= max_line) break; } return i; } int loadLabelName(const char *locationFilename, char *label[], int classNum) { printf("loadLabelName %sn", locationFilename); readLines(locationFilename, label, classNum); return 0; } static float CalculateOverlap(float xmin0, float ymin0, float xmax0, float ymax0, float xmin1, float ymin1, float xmax1, float ymax1) { float w = fmax(0.f, fmin(xmax0, xmax1) - fmax(xmin0, xmin1) + 1.0); float h = fmax(0.f, fmin(ymax0, ymax1) - fmax(ymin0, ymin1) + 1.0); float i = w * h; float u = (xmax0 - xmin0 + 1.0) * (ymax0 - ymin0 + 1.0) + (xmax1 - xmin1 + 1.0) * (ymax1 - ymin1 + 1.0) - i; return u <= 0.f ? 0.f : (i / u); } static int nms(int validCount, std::vector { for (int i = 0; i < validCount; ++i) { if (order == -1 || classIds != filterId) { continue; } int n = order; for (int j = i + 1; j < validCount; ++j) { int m = order[j]; if (m == -1 || classIds != filterId) { continue; } float xmin0 = outputLocations[n * 4 + 0]; float ymin0 = outputLocations[n * 4 + 1]; float xmax0 = outputLocations[n * 4 + 0] + outputLocations[n * 4 + 2]; float ymax0 = outputLocations[n * 4 + 1] + outputLocations[n * 4 + 3]; float xmin1 = outputLocations[m * 4 + 0]; float ymin1 = outputLocations[m * 4 + 1]; float xmax1 = outputLocations[m * 4 + 0] + outputLocations[m * 4 + 2]; float ymax1 = outputLocations[m * 4 + 1] + outputLocations[m * 4 + 3]; float iou = CalculateOverlap(xmin0, ymin0, xmax0, ymax0, xmin1, ymin1, xmax1, ymax1); if (iou > threshold) { order[j] = -1; } } } return 0; } static int quick_sort_indice_inverse( std::vector int left, int right, std::vector { float key; int key_index; int low = left; int high = right; if (left < right) { key_index = indices[left]; key = input[left]; while (low < high) { while (low < high && input[high] <= key) { high--; } input[low] = input[high]; indices[low] = indices[high]; while (low < high && input[low] >= key) { low++; } input[high] = input[low]; indices[high] = indices[low]; } input[low] = key; indices[low] = key_index; quick_sort_indice_inverse(input, left, low - 1, indices); quick_sort_indice_inverse(input, low + 1, right, indices); } return low; } static float sigmoid(float x) { return 1.0 / (1.0 + expf(-x)); } static float unsigmoid(float y) { return -1.0 * logf((1.0 / y) - 1.0); } inline static int32_t __clip(float val, float min, float max) { float f = val <= min ? min : (val >= max ? max : val); return f; } static uint8_t qnt_f32_to_affine(float f32, uint32_t zp, float scale) { float dst_val = (f32 / scale) + zp; uint8_t res = (uint8_t)__clip(dst_val, 0, 255); return res; } static float deqnt_affine_to_f32(uint8_t qnt, uint32_t zp, float scale) { return ((float)qnt - (float)zp) * scale; } static int process(uint8_t *input, int *anchor, int grid_h, int grid_w, int height, int width, int stride, std::vector float threshold, uint32_t zp, float scale) { int validCount = 0; int grid_len = grid_h * grid_w;//80*80, 40*40, 20*20. float thres = unsigmoid(threshold); uint8_t thres_u8 = qnt_f32_to_affine(thres, zp, scale);//量化 for (int a = 0; a < 3; a++) { for (int i = 0; i < grid_h; i++) { for (int j = 0; j < grid_w; j++) { /********************************************************************************************** 排列顺序是x y w h boxScore x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 idScore，然后保存的时候是先存 grid_h * grid_w个x，再存grid_h * grid_w个y,....,这样保存的。 而不是先存第一个x y w h boxScore x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 idScore， 再存第二个x y w h boxScore x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 idScore， **********************************************************************************************/ uint8_t box_confidence = input[(PROP_BOX_SIZE * a + 4) * grid_len + i * grid_w + j]; if (box_confidence >= thres_u8)//thres_u8经过反sigmoid和反量化操作了。 { int offset = (PROP_BOX_SIZE * a) * grid_len + i * grid_w + j;// uint8_t *in_ptr = input + offset; /*这边乘以2，减掉0.5是因为内部运算的时候做了这个处理，所以这里要反操作一下, deqnt_affine_to_f32是反量化，zp是量化时的zero point中心点，scale是量化时的尺度。*/ float box_x = sigmoid(deqnt_affine_to_f32(*in_ptr, zp, scale)) * 2.0 - 0.5; float box_y = sigmoid(deqnt_affine_to_f32(in_ptr[grid_len], zp, scale)) * 2.0 - 0.5; float box_w = sigmoid(deqnt_affine_to_f32(in_ptr[2 * grid_len], zp, scale)) * 2.0; float box_h = sigmoid(deqnt_affine_to_f32(in_ptr[3 * grid_len], zp, scale)) * 2.0; float x1 = (deqnt_affine_to_f32(in_ptr[5 * grid_len], zp, scale)); float y1 = (deqnt_affine_to_f32(in_ptr[6 * grid_len], zp, scale)); float x2 = (deqnt_affine_to_f32(in_ptr[7 * grid_len], zp, scale)); float y2 = (deqnt_affine_to_f32(in_ptr[8 * grid_len], zp, scale)); float x3 = (deqnt_affine_to_f32(in_ptr[9 * grid_len], zp, scale)); float y3 = (deqnt_affine_to_f32(in_ptr[10 * grid_len], zp, scale)); float x4 = (deqnt_affine_to_f32(in_ptr[11 * grid_len], zp, scale)); float y4 = (deqnt_affine_to_f32(in_ptr[12 * grid_len], zp, scale)); float x5 = (deqnt_affine_to_f32(in_ptr[13 * grid_len], zp, scale)); float y5 = (deqnt_affine_to_f32(in_ptr[14 * grid_len], zp, scale)); std::cout<<"landmark after deqnt_affine_to_f32:"< box_x = (box_x + j) * (float)stride;//这边的box_x是0-1的，相当于是偏移值，所以要加上j。乘以stride是映射回原图640. box_y = (box_y + i) * (float)stride;//这边的box_y是0-1的，相当于是偏移值，所以要加上i。乘以stride是映射回原图640. box_w = box_w * box_w * (float)anchor[a * 2]; box_h = box_h * box_h * (float)anchor[a * 2 + 1]; box_x -= (box_w / 2.0); box_y -= (box_h / 2.0); boxes.push_back(box_x); boxes.push_back(box_y); boxes.push_back(box_w); boxes.push_back(box_h); x1 = x1 * (float)anchor[a * 2] + j*(float)stride; y1 = y1 * (float)anchor[a * 2 + 1] + i*(float)stride; x2 = x2 * (float)anchor[a * 2] + j*(float)stride; y2 = y2 * (float)anchor[a * 2 + 1] + i*(float)stride; x3 = x3 * (float)anchor[a * 2] + j*(float)stride; y3 = y3 * (float)anchor[a * 2 + 1] + i*(float)stride; x4 = x4 * (float)anchor[a * 2] + j*(float)stride; y4 = y4 * (float)anchor[a * 2 + 1] + i*(float)stride; x5 = x5 * (float)anchor[a * 2] + j*(float)stride; y5 = y5 * (float)anchor[a * 2 + 1] + i*(float)stride; landMarks.push_back(x1); landMarks.push_back(y1); landMarks.push_back(x2); landMarks.push_back(y2); landMarks.push_back(x3); landMarks.push_back(y3); landMarks.push_back(x4); landMarks.push_back(y4); landMarks.push_back(x5); landMarks.push_back(y5); // printf("box_x=%.03f, box_y=%.03f, box_w=%.03f, box_h=%.03fn" // , box_x, box_y, box_w, box_h); uint8_t maxClassProbs = in_ptr[15 * grid_len]; int maxClassId = 0; for (int k = 0; k < OBJ_CLASS_NUM; ++k) { uint8_t prob = in_ptr[(15 + k) * grid_len];//这里是通过比较找到一个最高的得分和他的id。 if (prob > maxClassProbs) { maxClassId = k; maxClassProbs = prob; } } objProbs.push_back(sigmoid(deqnt_affine_to_f32(maxClassProbs, zp, scale))); classId.push_back(maxClassId); //boxes objProbs classId这三个变量往外返回的时候应该是每四个box对应一个objProb和一个classId。 //std::cout<<"maxClassProbs in post_process:::"< } } } } return validCount; } int post_process(uint8_t *input0, uint8_t *input1, uint8_t *input2, int model_in_h, int model_in_w, float conf_threshold, float nms_threshold, float scale_w, float scale_h, std::vector detect_result_group_t *group) { static int init = -1; if (init == -1) { int ret = 0; ret = loadLabelName(LABEL_NALE_TXT_PATH, labels, OBJ_CLASS_NUM); if (ret < 0) { return -1; } init = 0; } memset(group, 0, sizeof(detect_result_group_t)); std::vector std::vector std::vector std::vector // stride 8 int stride0 = 8; int grid_h0 = model_in_h / stride0;//80*80 int grid_w0 = model_in_w / stride0; int validCount0 = 0; validCount0 = process(input0, (int *)anchor0, grid_h0, grid_w0, model_in_h, model_in_w, stride0, filterBoxes, landMarks, objProbs, classId, conf_threshold, qnt_zps[0], qnt_scales[0]); // stride 16 int stride1 = 16; int grid_h1 = model_in_h / stride1;//40*40 int grid_w1 = model_in_w / stride1; int validCount1 = 0; validCount1 = process(input1, (int *)anchor1, grid_h1, grid_w1, model_in_h, model_in_w, stride1, filterBoxes, landMarks, objProbs, classId, conf_threshold, qnt_zps[1], qnt_scales[1]); // stride 32 int stride2 = 32; int grid_h2 = model_in_h / stride2;//20*20 int grid_w2 = model_in_w / stride2; int validCount2 = 0; validCount2 = process(input2, (int *)anchor2, grid_h2, grid_w2, model_in_h, model_in_w, stride2, filterBoxes, landMarks, objProbs, classId, conf_threshold, qnt_zps[2], qnt_scales[2]); int validCount = validCount0 + validCount1 + validCount2; printf("validCount=%dn", validCount); // no object detect if (validCount <= 0) { return 0; } std::vector for (int i = 0; i < validCount; ++i) { indexArray.push_back(i);//第i个是i，这是做什么用的。 } quick_sort_indice_inverse(objProbs, 0, validCount - 1, indexArray);//猜测这个函数是按照得分把index进行了排序。 std::set for (auto c : class_set)//c是int类型的。 { std::cout<<"c:::"< nms(validCount, filterBoxes, classId, indexArray, c, nms_threshold); } int last_count = 0; group->count = 0; /* box valid detect target */ for (int i = 0; i < validCount; ++i) { if (indexArray == -1 || i >= OBJ_NUMB_MAX_SIZE) { continue; } int n = indexArray; float x1 = filterBoxes[n * 4 + 0]; float y1 = filterBoxes[n * 4 + 1]; float x2 = x1 + filterBoxes[n * 4 + 2]; float y2 = y1 + filterBoxes[n * 4 + 3]; int id = classId[n]; float obj_conf = objProbs; float landmark_x1 = landMarks[n * 10 + 0]; float landmark_y1 = landMarks[n * 10 + 1]; float landmark_x2 = landMarks[n * 10 + 2]; float landmark_y2 = landMarks[n * 10 + 3]; float landmark_x3 = landMarks[n * 10 + 4]; float landmark_y3 = landMarks[n * 10 + 5]; float landmark_x4 = landMarks[n * 10 + 6]; float landmark_y4 = landMarks[n * 10 + 7]; float landmark_x5 = landMarks[n * 10 + 8]; float landmark_y5 = landMarks[n * 10 + 9]; // group->results[last_count].box.left = (int)(clamp(x1, 0, model_in_w) / scale_w); // group->results[last_count].box.top = (int)(clamp(y1, 0, model_in_h) / scale_h); // group->results[last_count].box.right = (int)(clamp(x2, 0, model_in_w) / scale_w); // group->results[last_count].box.bottom = (int)(clamp(y2, 0, model_in_h) / scale_h); group->results[last_count].box.left = (int)(clamp(x1, 0, model_in_w) );//clamp处理是否越界的。 group->results[last_count].box.top = (int)(clamp(y1, 0, model_in_h) ); group->results[last_count].box.right = (int)(clamp(x2, 0, model_in_w) ); group->results[last_count].box.bottom = (int)(clamp(y2, 0, model_in_h) ); group->results[last_count].landmark.push_back((int)(clamp(landmark_x1, 0, model_in_w))); group->results[last_count].landmark.push_back((int)(clamp(landmark_y1, 0, model_in_h))); group->results[last_count].landmark.push_back((int)(clamp(landmark_x2, 0, model_in_w))); group->results[last_count].landmark.push_back((int)(clamp(landmark_y2, 0, model_in_h))); group->results[last_count].landmark.push_back((int)(clamp(landmark_x3, 0, model_in_w))); group->results[last_count].landmark.push_back((int)(clamp(landmark_y3, 0, model_in_h))); group->results[last_count].landmark.push_back((int)(clamp(landmark_x4, 0, model_in_w))); group->results[last_count].landmark.push_back((int)(clamp(landmark_y4, 0, model_in_h))); group->results[last_count].landmark.push_back((int)(clamp(landmark_x5, 0, model_in_w))); group->results[last_count].landmark.push_back((int)(clamp(landmark_y5, 0, model_in_h))); group->results[last_count].prop = obj_conf; char *label = labels[id]; strncpy(group->results[last_count].name, label, OBJ_NAME_MAX_SIZE); // printf("result %2d: (%4d, %4d, %4d, %4d), %sn", i, group->results[last_count].box.left, group->results[last_count].box.top, // group->results[last_count].box.right, group->results[last_count].box.bottom, label); last_count++; } group->count = last_count; return 0; } int rknn_GetTop( float *pfProb, float *pfMaxProb, uint32_t *pMaxClass, uint32_t outputCount, uint32_t topNum) { uint32_t i, j; #define MAX_TOP_NUM 20 if (topNum > MAX_TOP_NUM) return 0; memset(pfMaxProb, 0, sizeof(float) * topNum); memset(pMaxClass, 0xff, sizeof(float) * topNum); for (j = 0; j < topNum; j++) { for (i = 0; i < outputCount; i++) { if ((i == *(pMaxClass + 0)) || (i == *(pMaxClass + 1)) || (i == *(pMaxClass + 2)) || (i == *(pMaxClass + 3)) || (i == *(pMaxClass + 4))) { continue; } if (pfProb > *(pfMaxProb + j)) { *(pfMaxProb + j) = pfProb; *(pMaxClass + j) = i; } } } return 1; } 4.3 main.cc // Copyright (c) 2021 by Rockchip Electronics Co., Ltd. All Rights Reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. /*------------------------------------------- Includes -------------------------------------------*/ #include #include #include #include #include #include #include #define _BASETSD_H #define STB_IMAGE_IMPLEMENTATION #include "stb/stb_image.h" #define STB_IMAGE_RESIZE_IMPLEMENTATION #include #undef cimg_display #define cimg_display 0 #include "CImg/CImg.h" #include "drm_func.h" #include "rga_func.h" #include "rknn_api.h" #include "postprocess.h" #include "async_client.h" #include "mqtt.h" using namespace cv; #define PERF_WITH_POST 1 #define REPVGG_LABEL_NALE_TXT_PATH "./model/repvgg_label.txt" #define logchw printf("function:%s,line:%d=========n", __FUNCTION__, __LINE__); using namespace std; const string DFLT_SERVER_ADDRESS { "tcp://192.168.1.10:1883" }; const string CLIENT_ID { "paho_cpp_async_publish" }; const string PERSIST_DIR { "./persist" }; const string TOPIC { "hello" }; const char* PAYLOAD1 = "Hello World!"; const char* PAYLOAD2 = "Hi there!"; const char* PAYLOAD3 = "Is anyone listening?"; const char* PAYLOAD4 = "Someone is always listening."; const char* LWT_PAYLOAD = "Last will and testament."; const int QOS = 1; const auto TIMEOUT = std::chrono::seconds(10); using namespace cimg_library; /*------------------------------------------- Functions -------------------------------------------*/ static void printRKNNTensor(rknn_tensor_attr *attr) { printf("index=%d name=%s n_dims=%d dims=[%d %d %d %d] n_elems=%d size=%d fmt=%d type=%d qnt_type=%d fl=%d zp=%d scale=%fn", attr->index, attr->name, attr->n_dims, attr->dims[3], attr->dims[2], attr->dims[1], attr->dims[0], attr->n_elems, attr->size, 0, attr->type, attr->qnt_type, attr->fl, attr->zp, attr->scale); } inline const char* get_type_string(rknn_tensor_type type) { switch(type) { case RKNN_TENSOR_FLOAT32: return "FP32"; case RKNN_TENSOR_FLOAT16: return "FP16"; case RKNN_TENSOR_INT8: return "INT8"; case RKNN_TENSOR_UINT8: return "UINT8"; case RKNN_TENSOR_INT16: return "INT16"; default: return "UNKNOW"; } } inline const char* get_qnt_type_string(rknn_tensor_qnt_type type) { switch(type) { case RKNN_TENSOR_QNT_NONE: return "NONE"; case RKNN_TENSOR_QNT_DFP: return "DFP"; case RKNN_TENSOR_QNT_AFFINE_ASYMMETRIC: return "AFFINE"; default: return "UNKNOW"; } } inline const char* get_format_string(rknn_tensor_format fmt) { switch(fmt) { case RKNN_TENSOR_NCHW: return "NCHW"; case RKNN_TENSOR_NHWC: return "NHWC"; default: return "UNKNOW"; } } static void dump_tensor_attr(rknn_tensor_attr *attr) { printf(" index=%d, name=%s, n_dims=%d, dims=[%d, %d, %d, %d], n_elems=%d, size=%d, fmt=%s, type=%s, qnt_type=%s, " "zp=%d, scale=%fn", attr->index, attr->name, attr->n_dims, attr->dims[3], attr->dims[2], attr->dims[1], attr->dims[0], attr->n_elems, attr->size, get_format_string(attr->fmt), get_type_string(attr->type), get_qnt_type_string(attr->qnt_type), attr->zp, attr->scale); } double __get_us(struct timeval t) { return (t.tv_sec * 1000000 + t.tv_usec); } static unsigned char *load_data(FILE *fp, size_t ofst, size_t sz) { unsigned char *data; int ret; data = NULL; if (NULL == fp) { return NULL; } ret = fseek(fp, ofst, SEEK_SET); if (ret != 0) { printf("blob seek failure.n"); return NULL; } data = (unsigned char *)malloc(sz); if (data == NULL) { printf("buffer malloc failure.n"); return NULL; } ret = fread(data, 1, sz, fp); return data; } static unsigned char *load_model(const char *filename, int *model_size) { FILE *fp; unsigned char *data; fp = fopen(filename, "rb"); if (NULL == fp) { printf("Open file %s failed.n", filename); return NULL; } fseek(fp, 0, SEEK_END); int size = ftell(fp); data = load_data(fp, 0, size); fclose(fp); *model_size = size; return data; } static unsigned char *load_repvgg_model(const char *filename, int *model_size) { FILE *fp = fopen(filename, "rb"); if (fp == nullptr) { printf("fopen %s fail!n", filename); return NULL; } fseek(fp, 0, SEEK_END); int model_len = ftell(fp); unsigned char *model = (unsigned char *)malloc(model_len); fseek(fp, 0, SEEK_SET); if (model_len != fread(model, 1, model_len, fp)) { printf("fread %s fail!n", filename); free(model); return NULL; } *model_size = model_len; if (fp) { fclose(fp); } return model; } void padding_resize(cv::InputArray src, cv::OutputArray dst, cv::Size size) { float padd_w = 0; float padd_h = 0; float r = std::min(float(size.width) / src.cols(), float(size.height) / src.rows()); int inside_w = round(src.cols() * r); int inside_h = round(src.rows() * r); padd_w = size.width - inside_w; padd_h = size.height - inside_h; cv::resize(src, dst, cv::Size(inside_w, inside_h)); padd_w = padd_w / 2; padd_h = padd_h / 2; //外层边框填充灰色 printf("gain:%f, padd_w:%f,padd_h:%f. in padding_resize============n", r, padd_w,padd_h); int top = int(round(padd_h - 0.1)); int bottom = int(round(padd_h + 0.1)); int left = int(round(padd_w - 0.1)); int right = int(round(padd_w + 0.1)); cv::copyMakeBorder(dst, dst, top, bottom, left, right, cv::BORDER_CONSTANT, cv::Scalar(114, 114, 114)); } cv::Mat read_image(const char *image_path, int w, int h, int & img_width, int & img_height, cv::Mat & img) { img = cv::imread(image_path); img_width = img.cols; img_height = img.rows; cv::Mat sample_resize, sample; padding_resize(img, sample_resize, cv::Size(w, h)); //cv::resize(img, sample_resize, cv::Size(w, h)); cv::cvtColor(sample_resize, sample, cv::COLOR_BGR2RGB); return sample; } /*------------------------------------------- Main Functions -------------------------------------------*/ int main(int argc, char **argv) { int status = 0; char *yolov5_model_name = NULL; rknn_context yolov5_ctx; unsigned int handle; size_t actual_size = 0; int yolov5_img_width = 0; int yolov5_img_height = 0; int img_channel = 0; const float nms_threshold = NMS_THRESH; const float box_conf_threshold = BOX_THRESH; struct timeval start_time, stop_time; int ret; rknn_context repvgg_ctx; int repvgg_model_len = 0; unsigned char *repvgg_model; int repvgg_img_width = 0; int repvgg_img_height = 0; //if (argc != 4) //{ //printf("Usage: %s //return -1; //} printf("post process config: box_conf_threshold = %.2f, nms_threshold = %.2fn", box_conf_threshold, nms_threshold); yolov5_model_name = (char *)argv[1]; char *image_name = argv[2]; string address = (argc > 3) ? string(argv[3]) : DFLT_SERVER_ADDRESS, clientID = (argc > 4) ? string(argv[4]) : CLIENT_ID; #ifdef MQTT //mqtt init cout << "Initializing for server '" << address << "'..." << endl; mqtt::async_client client(address, clientID, PERSIST_DIR); callback cb; client.set_callback(cb); auto connOpts = mqtt::connect_options_builder() .clean_session() .will(mqtt::message(TOPIC, LWT_PAYLOAD, QOS)) .finalize(); connOpts.set_keep_alive_interval(120 * 5); connOpts.set_clean_session(true); connOpts.set_automatic_reconnect(true); cout << "nConnecting..." << endl; mqtt::token_ptr conntok = client.connect(connOpts); cout << "Waiting for the connection..." << endl; conntok->wait(); cout << " .connect..OK" << endl; #endif /* Create the yolov5 neural network */ printf("===============================================Loading yolov5 mode=============================================n"); int model_data_size = 0; unsigned char *model_data = load_model(yolov5_model_name, &model_data_size); logchw ret = rknn_init(&yolov5_ctx, model_data, model_data_size, 0); if (ret < 0) { printf("rknn_init error ret=%dn", ret); return -1; } logchw rknn_sdk_version version; ret = rknn_query(yolov5_ctx, RKNN_QUERY_SDK_VERSION, &version, sizeof(rknn_sdk_version)); if (ret < 0) { printf("rknn_init error ret=%dn", ret); return -1; } printf("sdk version: %s driver version: %sn", version.api_version, version.drv_version); rknn_input_output_num io_num; ret = rknn_query(yolov5_ctx, RKNN_QUERY_IN_OUT_NUM, &io_num, sizeof(io_num)); if (ret < 0) { printf("rknn_init error ret=%dn", ret); return -1; } printf("model input num: %d, output num: %dn", io_num.n_input, io_num.n_output); rknn_tensor_attr input_attrs[io_num.n_input]; memset(input_attrs, 0, sizeof(input_attrs)); for (int i = 0; i < io_num.n_input; i++) { input_attrs.index = i; ret = rknn_query(yolov5_ctx, RKNN_QUERY_INPUT_ATTR, &(input_attrs), sizeof(rknn_tensor_attr)); if (ret < 0) { printf("rknn_init error ret=%dn", ret); return -1; } dump_tensor_attr(&(input_attrs)); } rknn_tensor_attr output_attrs[io_num.n_output]; memset(output_attrs, 0, sizeof(output_attrs)); for (int i = 0; i < io_num.n_output; i++) { output_attrs.index = i; ret = rknn_query(yolov5_ctx, RKNN_QUERY_OUTPUT_ATTR, &(output_attrs), sizeof(rknn_tensor_attr)); dump_tensor_attr(&(output_attrs)); if(output_attrs.qnt_type != RKNN_TENSOR_QNT_AFFINE_ASYMMETRIC || output_attrs.type != RKNN_TENSOR_UINT8) { fprintf(stderr,"The Demo required for a Affine asymmetric u8 quantized rknn model, but output quant type is %s, output data type is %sn", get_qnt_type_string(output_attrs.qnt_type),get_type_string(output_attrs.type)); return -1; } } int channel = 3; int yolov5_width = 0; int yolov5_height = 0; if (input_attrs[0].fmt == RKNN_TENSOR_NCHW)//这种格式时，排列是WHCN，我们的模型时这种格式。 { printf("model is NCHW input fmtn"); yolov5_width = input_attrs[0].dims[0]; yolov5_height = input_attrs[0].dims[1]; } else//RKNN_TENSOR_NHWC，这种格式时排列是CWHN。 { printf("model is NHWC input fmtn"); yolov5_width = input_attrs[0].dims[1]; yolov5_height = input_attrs[0].dims[2]; } printf("model input yolov5_height=%d, yolov5_width=%d, channel=%dn", yolov5_height, yolov5_width, channel); rknn_input yolov5_inputs[1]; memset(yolov5_inputs, 0, sizeof(yolov5_inputs)); yolov5_inputs[0].index = 0; yolov5_inputs[0].type = RKNN_TENSOR_UINT8; yolov5_inputs[0].size = yolov5_width * yolov5_height * channel; yolov5_inputs[0].fmt = RKNN_TENSOR_NHWC; //RKNN_TENSOR_NHWC;demo代码原来是NHWC,你改成了NCHW， yolov5_inputs[0].pass_through = 0; cv::Mat img; cv::Mat sample = read_image(image_name, yolov5_width, yolov5_height,yolov5_img_width,yolov5_img_height, img); yolov5_inputs[0].buf = sample.data; for(int i =0; i< sample.rows ;i++) { for(int j = 0; j < sample.cols; j++) { //std::cout<<"sample0:::"<<(double)sample.at //std::cout<<"sample1:::"<<(double)sample.at //std::cout<<"sample2::"<<(double)sample.at } } printf("yolov5_img_width:%dn, yolov5_img_height:%dn", yolov5_img_width,yolov5_img_height); gettimeofday(&start_time, NULL); rknn_inputs_set(yolov5_ctx, io_num.n_input, yolov5_inputs);//包含了颜色通道交换、归一化、量化、NHWC 转换成 NCHW 的过程 rknn_output yolov5_outputs[io_num.n_output]; memset(yolov5_outputs, 0, sizeof(yolov5_outputs)); for (int i = 0; i < io_num.n_output; i++) { yolov5_outputs.want_float = 0; } ret = rknn_run(yolov5_ctx, NULL); ret = rknn_outputs_get(yolov5_ctx, io_num.n_output, yolov5_outputs, NULL); //post process float scale_w = (float)yolov5_width / yolov5_img_width; float scale_h = (float)yolov5_height / yolov5_img_height; detect_result_group_t detect_result_group; std::vector std::vector for (int i = 0; i < io_num.n_output; ++i) { out_scales.push_back(output_attrs.scale); out_zps.push_back(output_attrs.zp); cout<<"out_scales["< printf("scale_w::::::::::%f,scale_h::::::%fn", scale_w, scale_h); post_process((uint8_t *)yolov5_outputs[0].buf, (uint8_t *)yolov5_outputs[1].buf, (uint8_t *)yolov5_outputs[2].buf, yolov5_height, yolov5_width, box_conf_threshold, nms_threshold, scale_w, scale_h, out_zps, out_scales, &detect_result_group); gettimeofday(&stop_time, NULL); printf("once run use %f msn", (__get_us(stop_time) - __get_us(start_time)) / 1000); // Draw Objects char text[256]; char resName[256]; char saveName[256]; const unsigned char blue[] = {0, 0, 255}; const unsigned char white[] = {255, 255, 255}; printf("detect_result_group.count::::%dn", detect_result_group.count); printf("yolov5_width:%d, yolov5_img_width:%d, yolov5_height:%d, yolov5_img_height:%d, after processn", yolov5_width, yolov5_img_width, yolov5_height, yolov5_img_height); //float gain = std::min(float(width) / img_width, float(height) / img_height); float gain = scale_w; if( scale_h < scale_w ) { gain = scale_h; } printf("gain:%fn", gain); float pad0 = (yolov5_width - yolov5_img_width * gain) / 2; float pad1 = (yolov5_height - yolov5_img_height * gain) / 2; std::vector std::vector for (int i = 0; i < detect_result_group.count; i++) { detect_result_t *det_result = &(detect_result_group.results); sprintf(text, "%s %.2f", det_result->name, det_result->prop); printf("%s @ (%d %d %d %d) %fn", det_result->name, det_result->box.left, det_result->box.top, det_result->box.right, det_result->box.bottom, det_result->prop); int x1 = det_result->box.left; int y1 = det_result->box.top; int x2 = det_result->box.right; int y2 = det_result->box.bottom; x1 -= pad0; y1 -= pad1; x2 -= pad0; y2 -= pad1; x1 /= gain; y1 /= gain; x2 /= gain; y2 /= gain; int landmark_x1 = (det_result->landmark[0] - pad0)/gain; int landmark_y1 = (det_result->landmark[1] - pad1)/gain; int landmark_x2 = (det_result->landmark[2] - pad0)/gain; int landmark_y2 = (det_result->landmark[3] - pad1)/gain; int landmark_x3 = (det_result->landmark[4] - pad0)/gain; int landmark_y3 = (det_result->landmark[5] - pad1)/gain; int landmark_x4 = (det_result->landmark[6] - pad0)/gain; int landmark_y4 = (det_result->landmark[7] - pad1)/gain; int landmark_x5 = (det_result->landmark[8] - pad0)/gain; int landmark_y5 = (det_result->landmark[9] - pad1)/gain; printf("det_result->box.left:%dn, det_result->box.top:%d,det_result->box.right:%d,det_result->box.bottom:%dn", det_result->box.left,det_result->box.top,det_result->box.right,det_result->box.bottom); printf("pad0:%f, pad1:%f, gain:%fn",pad0, pad1, gain); printf("x1:%d, y1:%d,x2:%d,y2:%dn",x1, y1, x2, y2); cv::rectangle(img, cv::Point(x1, y1), cv::Point(x2, y2), CV_RGB(0, 255, 0)); circle(img, cv::Point(landmark_x1, landmark_y1), 3,Scalar(255,0,0),-1); circle(img, cv::Point(landmark_x2, landmark_y2), 3,Scalar(255,0,0),-1); circle(img, cv::Point(landmark_x3, landmark_y3), 3,Scalar(255,0,0),-1); circle(img, cv::Point(landmark_x4, landmark_y4), 3,Scalar(255,0,0),-1); circle(img, cv::Point(landmark_x5, landmark_y5), 3,Scalar(0,255,0),-1); std::cout<<"landmark::====="< putText(img,text,Point(x1,y1-12),FONT_HERSHEY_SIMPLEX,1,Scalar(255,23,0),2,8); } cv::imwrite("./out.jpg", img); ret = rknn_outputs_release(yolov5_ctx, io_num.n_output, yolov5_outputs); #ifdef MQTT // Now try with a listener cout << "nSending algorithm result ..." << endl; action_listener listener; mqtt::message_ptr pubmsg = mqtt::make_message(TOPIC, repvggResult.c_str()); mqtt::delivery_token_ptr pubtok = client.publish(pubmsg, nullptr, listener); pubtok->wait(); cout << " ...OK" << endl; #endif #if 0 //yolov5 loop test for(int i = 0; i < 5; i++) { cv::Mat img; cv::Mat sample = read_image(image_name, yolov5_width, yolov5_height,yolov5_img_width,yolov5_img_height, img); yolov5_inputs[0].buf = sample.data; printf("yolov5_img_width:%dn, yolov5_img_height:%dn", yolov5_img_width,yolov5_img_height); gettimeofday(&start_time, NULL); rknn_inputs_set(yolov5_ctx, io_num.n_input, yolov5_inputs);//包含了颜色通道交换、归一化、量化、NHWC 转换成 NCHW 的过程。 rknn_output yolov5_outputs[io_num.n_output]; memset(yolov5_outputs, 0, sizeof(yolov5_outputs)); for (int i = 0; i < io_num.n_output; i++) { yolov5_outputs.want_float = 0; } ret = rknn_run(yolov5_ctx, NULL); ret = rknn_outputs_get(yolov5_ctx, io_num.n_output, yolov5_outputs, NULL); //post process float scale_w = (float)yolov5_width / yolov5_img_width; float scale_h = (float)yolov5_height / yolov5_img_height; detect_result_group_t detect_result_group; std::vector std::vector for (int i = 0; i < io_num.n_output; ++i) { out_scales.push_back(output_attrs.scale); out_zps.push_back(output_attrs.zp); } printf("scale_w::::::::::%f,scale_h::::::%fn", scale_w, scale_h); post_process((uint8_t *)yolov5_outputs[0].buf, (uint8_t *)yolov5_outputs[1].buf, (uint8_t *)yolov5_outputs[2].buf, yolov5_height, yolov5_width, box_conf_threshold, nms_threshold, scale_w, scale_h, out_zps, out_scales, &detect_result_group); ret = rknn_outputs_release(yolov5_ctx, io_num.n_output, yolov5_outputs); gettimeofday(&stop_time, NULL); printf("once run use %f msn", (__get_us(stop_time) - __get_us(start_time)) / 1000); } #endif // release ret = rknn_destroy(yolov5_ctx); if (model_data) { free(model_data); } return 0; } |
|
|
|
|
你正在撰写答案
如果你是对答案或其他答案精选点评或询问，请使用“评论”功能。
基于米尔瑞芯微RK3576核心板/开发板的人脸疲劳检测应用方案
569 浏览 0 评论
839 浏览 1 评论
737 浏览 1 评论
1957 浏览 1 评论
3204 浏览 1 评论
/6 
关注我们的微信
下载发烧友APP
德赢Vwin官网 观察
版权所有 © 湖南华秋数字科技有限公司
德赢Vwin官网 （电路图） 湘公网安备 43011202000918 号 电信与信息服务业务经营许可证：合字B2-20210191
工商网监
湘ICP备2023018690号
评论
淘帖
611
