手把手教你百度飛槳PP-YOLOE部署到瑞芯微RK3588_pp飛槳怎么安裝_布衣神棍的博客-CSDN博客手把手教你百度飛槳PP-YOLOE部署到瑞芯微RK3588https://blog.csdn.net/buyishengun/article/details/127653529開發(fā)板概況 — TB-RK3588x 0.1 文檔https://t.rock-chips.com/wiki/CN/tb-rk3588x/01%E5%BF%AB%E9%80%9F%E4%B8%8A%E6%89%8B.html瑞芯微RK3588等AI硬件NPU部署（paddle）_Vertira的博客-CSDN博客瑞芯微RK3588等AI硬件NPU部署https://blog.csdn.net/Vertira/article/details/127811686yolov5訓(xùn)練pt模型并轉(zhuǎn)換為rknn模型，部署在RK3588開發(fā)板上——從訓(xùn)練到部署全過程_rknn yolov5_Billy_zz的博客-CSDN博客本文實(shí)現(xiàn)了yolov5模型從訓(xùn)練pt模型，到轉(zhuǎn)換為rknn模型，最終部署在RK3588板子上使用NPU加速推理的過程。https://blog.csdn.net/m0_57315535/article/details/128250096?spm=1001.2101.3001.6650.5&utm_medium=distribute.pc_relevant.none-task-blog-2~default~CTRLIST~Rate-5-128250096-blog-126153227.pc_relevant_3mothn_strategy_and_data_recovery&depth_1-utm_source=distribute.pc_relevant.none-task-blog-2~default~CTRLIST~Rate-5-128250096-blog-126153227.pc_relevant_3mothn_strategy_and_data_recovery&utm_relevant_index=9rockchip-linux · GitCodeGitCode——開源代碼托管平臺，獨(dú)立第三方開源社區(qū)，Git/Github/Gitlabhttps://gitcode.net/mirrors/rockchip-linuxFirefly | 讓科技更簡單，讓生活更智能Firefly是天啟科技旗下的品牌，我們專注于開源智能硬件，物聯(lián)網(wǎng)，數(shù)字音頻產(chǎn)品的研發(fā)設(shè)計、生產(chǎn)和銷售，同時提供了智能硬件 產(chǎn)品的整體解決方案。Firefly產(chǎn)品包括行業(yè)主板，核心板，開源板等。全系列產(chǎn)品均是芯片原廠瑞芯微（Rockchip）推薦板卡，獲得原生SDK支持。核心板與行業(yè)主板廣泛應(yīng)用于人工智能，商業(yè)顯示，廣告一體機(jī)，智能POS，人臉識別終端，物聯(lián)網(wǎng)，智慧城市等領(lǐng)域。https://www.t-firefly.com/doc/download/106.html1. NPU使用 — Firefly WikiEC-R3588SPC采用 Rockchip RK3588S新一代旗艦級八核64位處理器，最大可配32GB大內(nèi)存；支持8K視頻編解碼；支持千兆網(wǎng)、；支持多種操作系統(tǒng)；可適用于ARM PC、邊緣計算、云服務(wù)器、智能NVR等領(lǐng)域https://wiki.t-firefly.com/zh_CN/EC-R3588SPC/usage_npu.html【FastDeploy + 瑞芯微】RV1126、RK3588全量化部署詳解_嗶哩嗶哩_bilibili【FastDeploy + 瑞芯微】RV1126、RK3588全量化部署詳解, 視頻播放量 618、彈幕量 0、點(diǎn)贊數(shù) 9、投硬幣枚數(shù) 4、收藏人數(shù) 24、轉(zhuǎn)發(fā)人數(shù) 6, 視頻作者 飛槳PaddlePaddle, 作者簡介 后廠村第一煉丹師<(￣︶￣)>，相關(guān)視頻：【瑞芯微NPU部署】官方Y(jié)OLOV5+Tengine推理引擎+RV1126嵌入式AI硬件，3588--新一代NPU的介紹及使用，瑞芯微RK3588 ARM PC解決方案，擺脫電腦，直接在RK3588平臺上推理，幫用戶部署 RKNN 模加速 AI 應(yīng)用的落地，瑞芯微RK3588高端平板解決方案，開箱首發(fā)！瑞芯微旗艦芯RK3588開發(fā)板，瑞芯微RK3588邊緣計算及AI應(yīng)用，瑞芯微RK3588智能車載360°全景環(huán)視，rv1126板子說明！，香蕉派開源社區(qū)完成瑞芯微RK3568/RK3588全國產(chǎn)化開發(fā)板硬件驗(yàn)證并運(yùn)行國產(chǎn)麒麟Linux系統(tǒng)https://www.bilibili.com/video/BV1Xg411x7MV/?spm_id_from=333.999.top_right_bar_window_history.content.click&vd_source=4aed82e35f26bb600bc5b46e65e25c22examples/vision/detection/paddledetection/rknpu2/README_CN.md · develop · mirrors / paddlepaddle / fastdeploy · GitCode??An Easy-to-use and Fast Deep Learning Model Deployment Toolkit for ??Cloud 📱Mobile andhttps://gitcode.net/mirrors/paddlepaddle/fastdeploy/-/blob/develop/examples/vision/detection/paddledetection/rknpu2/README_CN.mdyolov5訓(xùn)練并生成rknn模型以及3588平臺部署_rknn yolov5_新鑫信心的博客-CSDN博客瑞芯微RK3588上yolov5目標(biāo)檢測的部署。https://blog.csdn.net/m0_51714298/article/details/125916417

1.bfloat16的安裝

使用distutils構(gòu)建Python擴(kuò)展模塊(Building Python extension module with distutils)_電腦培訓(xùn)使用distutils構(gòu)建Python擴(kuò)展模塊(Building Python extension module with distutils),我正在使用distutils來構(gòu)建一個用C ++編寫的Python擴(kuò)展模塊。 我遇到的問題是，為了編譯https://www.656463.com/wenda/sydistutilsgjPythonkzmk_284https://www.cnblogs.com/goldsunshine/p/8872623.htmlhttps://www.cnblogs.com/goldsunshine/p/8872623.html用源碼編譯，直接sudo pip install 安裝有問題，源碼python setup.py install --user，有的時候sudo pip install 不好使，就使用--user這種方式，有的時候三方庫鏈接不到，python setup.py install不好使，可以試試python setup.py sdist，記住distutils只是一種打包方式。此外，升級的gcc5.4在源碼編譯時可能有問題，加上

module.extra_compile_args = ['--std=c++0x']

2.rknpu

?250幀，640x640

?訓(xùn)練框架自帶的一些量化操作可以被rknn讀取

rknpu即為上面的rknn runtime.

rk3588的cpu比較強(qiáng)，其實(shí)在PC或者板子上編譯都可以。

dmesg | grep Galcore 查看版本

?wget/adb/ssh推送都是可以的

rk3588對輸入有優(yōu)化

3.paddledetection中的pp-human在rk3588上部署

rknn的部署其實(shí)有兩部分，第一部分是在linux系統(tǒng)上開發(fā)量化和rknn模型，用到的是rknn-toolkit2，在板子上python調(diào)用，pc端是必須要裝的，因?yàn)榻o的rknn-toolkit2是用來轉(zhuǎn)模型的，如果不用fastdeploy這樣的框架的話。可以使用rknn-toolkit2中rknn-lite來部署，可以調(diào)用npu接口，如果對速度有更高要求，則需要使用rknpu來進(jìn)行c++側(cè)的部署，c++優(yōu)化了預(yù)處理等操作，事實(shí)上，預(yù)處理是很耗時的。

3.1 paddle2rknn

paddle提供的權(quán)重已經(jīng)是export_model之后的模型了，因此在export_model.py之前通過exclude_nms=True不可行。

下面這個鏈接可行
examples/vision/detection/paddledetection/rknpu2/README_CN.md · develop · mirrors / paddlepaddle / fastdeploy · GitCode??An Easy-to-use and Fast Deep Learning Model Deployment Toolkit for ??Cloud 📱Mobile andhttps://gitcode.net/mirrors/paddlepaddle/fastdeploy/-/blob/develop/examples/vision/detection/paddledetection/rknpu2/README_CN.md

代碼示例：

paddle2onnx --model_dir /home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/paddle_ori/mot_ppyoloe_l_36e_pipeline \ --model_filename /home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/paddle_ori/mot_ppyoloe_l_36e_pipeline/model.pdmodel \ --params_filename /home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/paddle_ori/mot_ppyoloe_l_36e_pipeline/model.pdiparams \ --save_file /home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/ppyoloeonnx/mot_ppyoloe_l_36e_pipeline.onnx \ --enable_dev_version True --opset_version 12 --enable_onnx_checker True python -m paddle2onnx.optimize --input_model /home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/ppyoloeonnx/mot_ppyoloe_l_36e_pipeline.onnx \ --output_model /home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/ppyoloeonnx/mot_ppyoloe_l_36e_pipeline_opt.onnx \ --input_shape_dict "{'image':[1,3,640,640], 'scale_factor':[1,2]}" python export_rknn.py --config_path /home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/ppyoloeonnx/mot_ppyoloe_opt.yaml --target_platform rk3588

上面這種方式是通過rknn來進(jìn)行onnx靜態(tài)圖的裁剪，避免了在onnx和paddle上面進(jìn)行裁剪。

3.2 paddleinference->rknn推理

剪裁完節(jié)點(diǎn)之后輸出是[(8400,4),8400]的array，第一維是兩個點(diǎn)坐標(biāo)有8400，后面是對應(yīng)的置信度有8400個，此時解耦頭輸出的結(jié)果。

pphuman部署的是ppyoloe模型，其實(shí)是跟蹤的模型，但是第一部分還是ppyoloe，后面加上卡爾曼濾波和匈牙利算法，所以核心還是ppyoloe在rk3588的部署。

此處寫了兩個版本的nms去做后處理，發(fā)現(xiàn)都輸出的框都有問題，即便是沒有量化，輸出的框還是有問題。

import os import urllib import traceback import time import sys import yaml import numpy as np import cv2 from rknn.api import RKNNONNX_MODEL = 'yolov5s.onnx' RKNN_MODEL = '/home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/ppyoloeonnx/mot_ppyoloe_l_36e_pipeline_opt_rk3588_quantized.rknn' IMG_PATH = './test.png' DATASET = './dataset.txt'QUANTIZE_ON = TrueOBJ_THRESH = 0.25 NMS_THRESH = 0.45 IMG_SIZE = 640CLASSES = ["player"]def draw_results(result, image, draw_thresh=0.5): # plt.figure(figsize=(10, 10)) # im = imread(filename) # plt.imshow(im) # currentAxis=plt.gca() # colors = ['r', 'g', 'b', 'k', 'y', 'pink', 'purple']for item in result:top, left, right, bottom= item[2:6]label = int(item[0])score = item[1]name = CLASSES[label]if item[1] > draw_thresh:top = int(top)left = int(left)right = int(right)bottom = int(bottom)cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 2)cv2.putText(image, '{0} {1:.2f}'.format(name, score),(top, left - 6),cv2.FONT_HERSHEY_SIMPLEX,0.6, (0, 0, 255), 2) # draw_rectangle(currentAxis, box, edgecolor = colors[label]) # plt.text(box[0], box[1], name, fontsize=12, color=colors[label]) # plt.savefig('/home/aistudio/external-libraries/PaddleDection/outout_img/output_pic.png')# plt.show()def draw(image, boxes, scores, classes):"""Draw the boxes on the image.# Argument:image: original image.boxes: ndarray, boxes of objects.classes: ndarray, classes of objects.scores: ndarray, scores of objects.all_classes: all classes name."""for box, score, cl in zip(boxes, scores, classes):top, left, right, bottom = boxprint('class: {}, score: {}'.format(CLASSES[cl], score))print('box coordinate left,top,right,down: [{}, {}, {}, {}]'.format(top, left, right, bottom))top = int(top)left = int(left)right = int(right)bottom = int(bottom)cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 2)cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score),(top, left - 6),cv2.FONT_HERSHEY_SIMPLEX,0.6, (0, 0, 255), 2)# 計算IoU，矩形框的坐標(biāo)形式為xyxy，這個函數(shù)會被保存在box_utils.py文件中 def box_iou_xyxy(box1, box2):# 獲取box1左上角和右下角的坐標(biāo)x1min, y1min, x1max, y1max = box1[0], box1[1], box1[2], box1[3]# 計算box1的面積s1 = (y1max - y1min + 1.) * (x1max - x1min + 1.)# 獲取box2左上角和右下角的坐標(biāo)x2min, y2min, x2max, y2max = box2[0], box2[1], box2[2], box2[3]# 計算box2的面積s2 = (y2max - y2min + 1.) * (x2max - x2min + 1.)# 計算相交矩形框的坐標(biāo)xmin = np.maximum(x1min, x2min)ymin = np.maximum(y1min, y2min)xmax = np.minimum(x1max, x2max)ymax = np.minimum(y1max, y2max)# 計算相交矩形行的高度、寬度、面積inter_h = np.maximum(ymax - ymin + 1., 0.)inter_w = np.maximum(xmax - xmin + 1., 0.)intersection = inter_h * inter_w# 計算相并面積union = s1 + s2 - intersection# 計算交并比iou = intersection / unionreturn ioudef nms(bboxes, scores, score_thresh, nms_thresh, pre_nms_topk):"""nms"""inds = np.argsort(scores)inds = inds[::-1]inds = inds[:pre_nms_topk]keep_inds = []while(len(inds) > 0):cur_ind = inds[0]cur_score = scores[cur_ind]# if score of the box is less than score_thresh, just drop itif cur_score < score_thresh:breakkeep = Truefor ind in keep_inds:current_box = bboxes[cur_ind]remain_box = bboxes[ind]# import pdb;pdb.set_trace()iou = box_iou_xyxy(current_box, remain_box)if iou > nms_thresh:keep = Falsebreakif keep:keep_inds.append(cur_ind)inds = inds[1:]return np.array(keep_inds)def multiclass_nms(bboxes, scores, score_thresh=0.05, nms_thresh=0.5, pre_nms_topk=1000, pos_nms_topk=100):"""This is for multiclass_nms"""batch_size = bboxes.shape[0]class_num = scores.shape[1]rets = []for i in range(batch_size):bboxes_i = bboxes[i]scores_i = scores[i]ret = []for c in range(class_num):scores_i_c = scores_i[c]keep_inds = nms(bboxes_i, scores_i_c, score_thresh, nms_thresh, pre_nms_topk)if len(keep_inds) < 1:continuekeep_bboxes = bboxes_i[keep_inds]keep_scores = scores_i_c[keep_inds]keep_results = np.zeros([keep_scores.shape[0], 6])keep_results[:, 0] = ckeep_results[:, 1] = keep_scores[:]keep_results[:, 2:6] = keep_bboxes[:, :]ret.append(keep_results)if len(ret) < 1:rets.append(ret)continueret_i = np.concatenate(ret, axis=0)scores_i = ret_i[:, 1]if len(scores_i) > pos_nms_topk:inds = np.argsort(scores_i)[::-1]inds = inds[:pos_nms_topk]ret_i = ret_i[inds]rets.append(ret_i)return retsdef nms_boxes(boxes, scores):"""Suppress non-maximal boxes.# Argumentsboxes: ndarray, boxes of objects.scores: ndarray, scores of objects.# Returnskeep: ndarray, index of effective boxes."""x = boxes[:, 0]y = boxes[:, 1]w = boxes[:, 2] - boxes[:, 0]h = boxes[:, 3] - boxes[:, 1]areas = w * horder = 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 * h1ovr = inter / (areas[i] + areas[order[1:]] - inter)inds = np.where(ovr <= NMS_THRESH)[0]order = order[inds + 1]keep = np.array(keep)return keepdef ppyolo_nms(pred_bboxes, pred_scores): # pred_bboxes:(1,8400,1) pred_scores:(1,1,8400) boxes = pred_bboxes.reshape(-1,4)box_class_probs = pred_scores.reshape(pred_scores.shape[-1]*pred_scores.shape[0],-1)box_confidences = np.ones(boxes.shape[0]).reshape(-1,)_box_pos = np.where(box_confidences >= OBJ_THRESH)boxes = boxes[_box_pos]box_confidences = box_confidences[_box_pos]box_class_probs = box_class_probs[_box_pos]class_max_score = np.max(box_class_probs, axis=-1)classes = np.argmax(box_class_probs, axis=-1)_class_pos = np.where(class_max_score >= OBJ_THRESH)boxes = boxes[_class_pos]classes = classes[_class_pos]scores = (class_max_score* box_confidences)[_class_pos] # boxes = np.concatenate(boxes) # classes = np.concatenate(classes) # scores = np.concatenate(scores)nboxes, nclasses, nscores = [], [], []for c in set(classes):inds = np.where(classes == c)b = boxes[inds]c = classes[inds]s = scores[inds]keep = nms_boxes(b, s)nboxes.append(b[keep])nclasses.append(c[keep])nscores.append(s[keep])if not nclasses and not nscores:return None, None, Noneboxes = np.concatenate(nboxes)classes = np.concatenate(nclasses)scores = np.concatenate(nscores)return boxes, classes, scoresdef letterbox(im, new_shape=(640, 640), color=(0, 0, 0)):# Resize and pad image while meeting stride-multiple constraintsshape = im.shape[:2] # current shape [height, width]if isinstance(new_shape, int):new_shape = (new_shape, new_shape)# Scale ratio (new / old)r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])# Compute paddingratio = r, r # width, height ratiosnew_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh paddingdw /= 2 # divide padding into 2 sidesdh /= 2if shape[::-1] != new_unpad: # resizeim = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))left, right = int(round(dw - 0.1)), int(round(dw + 0.1))im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add borderreturn im, ratio, (dw, dh)if __name__ == '__main__': with open("/home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/ppyoloeonnx/mot_ppyoloe_opt.yaml") as file:file_data = file.read()yaml_config = yaml.safe_load(file_data)print(yaml_config)model = RKNN(True)# Configmean_values = yaml_config["mean"]std_values = yaml_config["std"]model.config(mean_values=mean_values,std_values=std_values,target_platform="rk3588")# Load ONNX modelif yaml_config["outputs_nodes"] is None:ret = model.load_onnx(model=yaml_config["model_path"])else:ret = model.load_onnx(model=yaml_config["model_path"],outputs=yaml_config["outputs_nodes"])assert ret == 0, "Load model failed!"# Build modelret = model.build(do_quantization=yaml_config["do_quantization"],dataset=yaml_config["dataset"])assert ret == 0, "Build model failed!"# Init Runtimeret = model.init_runtime()assert ret == 0, "Init runtime environment failed!"# Exportif not os.path.exists(yaml_config["output_folder"]):os.mkdir(yaml_config["output_folder"])model_base_name = os.path.basename(yaml_config["model_path"]).split(".")[0] # model_device_name = config.target_platform.lower()model_device_name = 'rk3588'if yaml_config["do_quantization"]:model_save_name = model_base_name + "_" + model_device_name + "_quantized" + ".rknn"else:model_save_name = model_base_name + "_" + model_device_name + "_unquantized" + ".rknn"ret = model.export_rknn(os.path.join(yaml_config["output_folder"], model_save_name))assert ret == 0, "Export rknn model failed!"print("Export OK!")# Set inputsimg = cv2.imread(IMG_PATH)img, ratio, (dw, dh) = letterbox(img, new_shape=(IMG_SIZE, IMG_SIZE))img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)img = cv2.resize(img, (IMG_SIZE, IMG_SIZE))# Inferenceprint('--> Running model')outputs = model.inference(inputs=[img])print('done') # import pdb;pdb.set_trace() # pred_bboxes = outputs[0] # pred_scores = outputs[1] # bbox_pred = multiclass_nms(pred_bboxes,pred_scores) # bbox_pred = bbox_pred[0]# import pdb;pdb.set_trace() # img_1 = cv2.cvtColor(img, cv2.COLOR_RGB2BGR) # if bbox_pred is not None: # draw_results(bbox_pred, img_1, draw_thresh=0.1) # cv2.imwrite("1.png",img_1) # draw_results(result, image, draw_thresh=0.5):import pdb;pdb.set_trace()pred_bboxes = outputs[0]pred_scores = outputs[1] # pred_bboxes = pred_bboxes.numpy() # pred_scores = pred_scores.numpy()boxes, classes, scores = ppyolo_nms(pred_bboxes,pred_scores)img_1 = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)if boxes is not None:draw(img_1, boxes, scores, classes)cv2.imwrite("1.png",img_1)rknn.release()

4.fastdeploy部署

fastdeploy可以部署在rk3588上，比較友好的其還有python的接口，可以做快速的嘗試，還是用ppyoloe跑檢測，后面在加上跟蹤的代碼，fastdeploy的python接口在rk3588上默認(rèn)是onnxruntime的推理后端，python似乎是不支持rknpu2，沒跑通。

''' @Time : 2023/2/27 14:37 @Author : leeguandon@gmail.com ''' import os import fastdeploy as fd import cv2 import numpy as np from ocsort_tracker import OCSORTTracker from collections import defaultdict from pathlib import Path import copyOBJ_THRESH = 0.25 NMS_THRESH = 0.45 CLASSES = ["player"] visual = Truedef draw(image, boxes, scores, classes):"""Draw the boxes on the image.# Argument:image: original image.boxes: ndarray, boxes of objects.classes: ndarray, classes of objects.scores: ndarray, scores of objects.all_classes: all classes name."""for box, score, cl in zip(boxes, scores, classes):top, left, right, bottom = boxprint('class: {}, score: {}'.format(CLASSES[cl], score))print('box coordinate left,top,right,down: [{}, {}, {}, {}]'.format(top, left, right, bottom))top = int(top)left = int(left)right = int(right)bottom = int(bottom)cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 2)cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score),(top, left - 6),cv2.FONT_HERSHEY_SIMPLEX,0.6, (0, 0, 255), 2)def nms_boxes(boxes, scores):"""Suppress non-maximal boxes.# Argumentsboxes: ndarray, boxes of objects.scores: ndarray, scores of objects.# Returnskeep: ndarray, index of effective boxes."""x = boxes[:, 0]y = boxes[:, 1]w = boxes[:, 2] - boxes[:, 0]h = boxes[:, 3] - boxes[:, 1]areas = w * horder = 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 * h1ovr = inter / (areas[i] + areas[order[1:]] - inter)inds = np.where(ovr <= NMS_THRESH)[0]order = order[inds + 1]keep = np.array(keep)return keepdef ppyolo_nms(pred_bboxes, pred_scores):boxes = pred_bboxes.reshape(-1, 4)box_confidences = np.ones(pred_bboxes.shape[0]).reshape(-1, )box_class_probs = pred_scores.reshape(pred_scores.shape[-1], -1)_box_pos = np.where(box_confidences >= OBJ_THRESH)boxes = boxes[_box_pos]box_confidences = box_confidences[_box_pos]box_class_probs = box_class_probs[_box_pos]class_max_score = np.max(box_class_probs, axis=-1)classes = np.argmax(box_class_probs, axis=-1)_class_pos = np.where(class_max_score >= OBJ_THRESH)boxes = boxes[_class_pos]classes = classes[_class_pos]scores = (class_max_score * box_confidences)[_class_pos]nboxes, nclasses, nscores = [], [], []for c in set(classes):inds = np.where(classes == c)b = boxes[inds]c = classes[inds]s = scores[inds]keep = nms_boxes(b, s)nboxes.append(b[keep])nclasses.append(c[keep])nscores.append(s[keep])if not nclasses and not nscores:return None, None, Noneboxes = np.concatenate(nboxes)classes = np.concatenate(nclasses)scores = np.concatenate(nscores)return boxes, classes, scoresdef postprocess(boxes, classes, scores):nboxes = []if len(boxes) > 0:for i, box in enumerate(boxes):classes_scores = np.append(classes[i], scores[i])boxes_ = np.append(classes_scores, box)nboxes.append(boxes_.tolist())result = {"boxes": np.array([boxes_]), 'boxes_num': np.array([len(boxes)])}return resultclass SDE_Detector(object):def __init__(self):use_byte = Falsedet_thresh = 0.4max_age = 30min_hits = 3iou_threshold = 0.3delta_t = 3inertia = 0.2min_box_area = 0vertical_ratio = 0self.tracker = OCSORTTracker(det_thresh=det_thresh,max_age=max_age,min_hits=min_hits,iou_threshold=iou_threshold,delta_t=delta_t,inertia=inertia,min_box_area=min_box_area,vertical_ratio=vertical_ratio,use_byte=use_byte)def tracking(self, det_results):pred_dets = det_results['boxes']pred_embs = det_results.get('embeddings', None)online_targets = self.tracker.update(pred_dets, pred_embs)online_tlwhs = defaultdict(list)online_scores = defaultdict(list)online_ids = defaultdict(list)for t in online_targets:tlwh = [t[0], t[1], t[2] - t[0], t[3] - t[1]] # top,left,w,htscore = float(t[4])tid = int(t[5])if tlwh[2] * tlwh[3] <= self.tracker.min_box_area: continueif self.tracker.vertical_ratio > 0 and tlwh[2] / tlwh[3] > self.tracker.vertical_ratio:continueif tlwh[2] * tlwh[3] > 0:online_tlwhs[0].append(tlwh)online_ids[0].append(tid)online_scores[0].append(tscore)tracking_outs = {'online_tlwhs': online_tlwhs, # 坐標(biāo)'online_scores': online_scores, # >0.4'online_ids': online_ids, # [10,9,8,7,6,5,4,3,2,1]}return tracking_outsdef get_color(idx):idx = idx * 3color = ((37 * idx) % 255, (17 * idx) % 255, (29 * idx) % 255)return colordef plot_tracking_dict(image,num_classes,tlwhs_dict,obj_ids_dict,scores_dict,frame_id=0,fps=0.,ids2names=[]):im = np.ascontiguousarray(np.copy(image)) # shape：480,854,3im_h, im_w = im.shape[:2]text_scale = max(0.5, image.shape[1] / 3000.)text_thickness = 2line_thickness = max(1, int(image.shape[1] / 500.))for cls_id in range(num_classes):tlwhs = tlwhs_dict[cls_id]obj_ids = obj_ids_dict[cls_id]scores = scores_dict[cls_id]cv2.putText(im,'frame: %d fps: %.2f num: %d' % (frame_id, fps, len(tlwhs)),(0, int(15 * text_scale) + 5),cv2.FONT_ITALIC,text_scale, (0, 0, 255),thickness=text_thickness)record_id = set()for i, tlwh in enumerate(tlwhs):x1, y1, w, h = tlwhintbox = tuple(map(int, (x1, y1, x1 + w, y1 + h)))center = tuple(map(int, (x1 + w / 2., y1 + h / 2.)))obj_id = int(obj_ids[i])id_text = '{}'.format(int(obj_id))if ids2names != []:id_text = '{}_{}'.format(ids2names[cls_id], id_text)else:id_text = 'class{}_{}'.format(cls_id, id_text)_line_thickness = 1 if obj_id <= 0 else line_thicknessin_region = Falsecolor = get_color(abs(obj_id)) if in_region == False else (0, 0,255)cv2.rectangle(im,intbox[0:2],intbox[2:4],color=color,thickness=line_thickness)cv2.putText(im,id_text, (intbox[0], intbox[1] - 25),cv2.FONT_ITALIC,text_scale,color,thickness=text_thickness)return imoption = fd.RuntimeOption() # option.use_cpu() # option.use_openvino_backend() # 一行命令切換使用 OpenVINO部署model = fd.vision.detection.PPYOLOE("/home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/paddle_ori/mot_ppyoloe_l_36e_pipeline/model.pdmodel","/home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/paddle_ori/mot_ppyoloe_l_36e_pipeline/model.pdiparams" , "/home/sniss/local_disk/rknn-toolkit2-master/examples/onnx/ppyoloe/paddle_ori/mot_ppyoloe_l_36e_pipeline/infer_cfg.yml")tracker = SDE_Detector() mot_results = []video_file = "kitch.mp4" output_dir = "results" capture = cv2.VideoCapture(video_file)# Get Video info : resolution, fps, frame count width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH)) height = int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT)) fps = int(capture.get(cv2.CAP_PROP_FPS)) frame_count = int(capture.get(cv2.CAP_PROP_FRAME_COUNT)) print("video fps: %d, frame_count: %d" % (fps, frame_count))video_out_name = Path(video_file).stem if not os.path.exists(output_dir):os.makedirs(output_dir) out_path = os.path.join(output_dir, video_out_name + ".mp4") fourcc = cv2.VideoWriter_fourcc(*'mp4v') writer = cv2.VideoWriter(out_path, fourcc, fps, (width, height))frame_id = 0while (1):# frame_id = 0if frame_id % 10 == 0:print('frame id: {}'.format(frame_id))ret, frame = capture.read()if not ret:break# img = cv2.imread("test.png")img = frameresult = model.predict(copy.deepcopy(img))pred_bboxes = np.array(result.boxes)pred_scores = np.array(result.scores)boxes, classes, scores = ppyolo_nms(pred_bboxes, pred_scores)# boxes# array([[ 618.63458252, 172.54750061, 1023.77459717, 781.89233398]])# classes# array([0])# scores# array([0.95259225])det_result = postprocess(boxes, classes, scores)tracking_outs = tracker.tracking(det_result)online_tlwhs = tracking_outs['online_tlwhs']online_scores = tracking_outs['online_scores']online_ids = tracking_outs['online_ids']mot_results.append([online_tlwhs, online_scores, online_ids])if visual:im = plot_tracking_dict(frame,1,online_tlwhs,online_ids,online_scores,frame_id=frame_id,ids2names=CLASSES)cv2.imwrite(os.path.join(output_dir, '{:05d}.jpg'.format(frame_id)), im)frame_id += 1writer.write(im)# img_1 = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)# if boxes is not None:# draw(img_1, boxes, scores, classes)## img_2 = cv2.cvtColor(img_1, cv2.COLOR_BGR2RGB)# cv2.imwrite("fastdeploy_1.png", img_2)writer.release()

fastdeploy的c++版本嘗試、

5.rknn-lite2 npu的python接口

用rknn-lite2可以調(diào)用rk3588的npu，速度相當(dāng)不錯。

總結(jié)：整體來說，在pc端可以使用rknn-toolkit2做rknn模型的轉(zhuǎn)換，這一步最關(guān)鍵的就是.rknn模型的獲取，rknpu中雖然也有onnx2rknn，但是想在pc端用python做測試，還是走這個路子，有了rknn模型之后，可以用rknn-lite2在rk3588上推理，或者使用rknpu的c++接口，rknpu的c++接口高效處理了數(shù)據(jù)預(yù)處理，這一步其實(shí)很耗時間的，速度會更快點(diǎn)。

總結(jié)

以上是生活随笔為你收集整理的pp-human在rk3588上部署的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯，歡迎將生活随笔推薦給好友。