当前位置: 首页 > news >正文

OpenMMlab导出yolov3模型并用onnxruntime和tensorrt推理

导出onnx文件

直接使用脚本

import torch
from mmdet.apis import init_detector, inference_detectorconfig_file = './configs/yolo/yolov3_mobilenetv2_8xb24-ms-416-300e_coco.py'
checkpoint_file = 'yolov3_mobilenetv2_mstrain-416_300e_coco_20210718_010823-f68a07b3.pth'
model = init_detector(config_file, checkpoint_file, device='cpu')  # or device='cuda:0'
torch.onnx.export(model, (torch.zeros(1, 3, 416, 416),), "out.onnx", opset_version=11)

导出的onnx结构如下:
在这里插入图片描述
输出是包含三个不同层级检测头的输出。若需要合并检测结果,需要修改脚本如下:

import torch
from itertools import repeat
from mmdet.apis import init_detector, inference_detectorconfig_file = './configs/yolo/yolov3_mobilenetv2_8xb24-ms-416-300e_coco.py'
checkpoint_file = 'yolov3_mobilenetv2_mstrain-416_300e_coco_20210718_010823-f68a07b3.pth'
model = init_detector(config_file, checkpoint_file, device='cpu')  # or device='cuda:0'class YOLOV3(torch.nn.Module):def __init__(self):super().__init__()self.model = init_detector(config_file, checkpoint_file, device='cpu')self.class_num = 80self.base_sizes = [[(116, 90), (156, 198), (373, 326)], [(30, 61), (62, 45), (59, 119)], [(10, 13), (16, 30), (33, 23)]]self.stride = [32, 16, 8]self.strides = [tuple(repeat(x, 2)) for x in self.stride]self.centers = [(x[0] / 2., x[1] / 2.) for x in self.strides]self.base_anchors=self.gen_base_anchors()def gen_base_anchors(self):multi_level_base_anchors = []for i, base_sizes_per_level in enumerate(self.base_sizes):center = self.centers[i]x_center, y_center = centerbase_anchors = []for base_size in base_sizes_per_level:w, h = base_sizebase_anchor = torch.Tensor([x_center - 0.5 * w, y_center - 0.5 * h, x_center + 0.5 * w, y_center + 0.5 * h])base_anchors.append(base_anchor)base_anchors = torch.stack(base_anchors, dim=0)multi_level_base_anchors.append(base_anchors)return multi_level_base_anchorsdef _meshgrid(self, x, y):xx = x.repeat(y.shape[0])yy = y.view(-1, 1).repeat(1, x.shape[0]).view(-1)return xx, yydef grid_priors(self, featmap_sizes):multi_level_anchors = []for i in range(len(featmap_sizes)):base_anchors = self.base_anchors[i]feat_h, feat_w = featmap_sizes[i]stride_w, stride_h = self.strides[i]shift_x = torch.arange(0, feat_w) * stride_wshift_y = torch.arange(0, feat_h) * stride_hshift_xx, shift_yy = self._meshgrid(shift_x, shift_y)shifts = torch.stack([shift_xx, shift_yy, shift_xx, shift_yy], dim=-1)anchors = base_anchors[None, :, :] + shifts[:, None, :]anchors = anchors.view(-1, 4)           multi_level_anchors.append(anchors)return multi_level_anchorsdef decode(self, bboxes, pred_bboxes, stride):xy_centers = (bboxes[..., :2] + bboxes[..., 2:]) * 0.5 + (pred_bboxes[..., :2] - 0.5) * stridewhs = (bboxes[..., 2:] - bboxes[..., :2]) * 0.5 * pred_bboxes[..., 2:].exp()decoded_bboxes = torch.stack((xy_centers[..., 0] - whs[..., 0], xy_centers[..., 1] - whs[..., 1], xy_centers[..., 0] + whs[..., 0], xy_centers[..., 1] + whs[..., 1]), dim=-1)return decoded_bboxesdef forward(self, x):x = self.model.backbone(x)x = self.model.neck(x)pred_maps = self.model.bbox_head(x)flatten_preds = []flatten_strides = []for pred, stride in zip(pred_maps[0], self.stride):pred = pred.permute(0, 2, 3, 1).reshape(1, -1, 5+self.class_num)pred[..., :2] = pred[..., :2].sigmoid()flatten_preds.append(pred)flatten_strides.append(pred.new_tensor(stride).expand(pred.size(1)))flatten_preds = torch.cat(flatten_preds, dim=1)flatten_bbox_preds = flatten_preds[..., :4]  flatten_objectness = flatten_preds[..., 4].sigmoid()flatten_preds[..., 4] = flatten_objectnessflatten_cls_scores = flatten_preds[..., 5:].sigmoid()flatten_preds[..., 5:] = flatten_cls_scoresfeatmap_sizes = [pred_map.shape[-2:] for pred_map in pred_maps[0]]mlvl_anchors = self.grid_priors(featmap_sizes)flatten_anchors = torch.cat(mlvl_anchors)flatten_strides = torch.cat(flatten_strides)flatten_bboxes = self.decode(flatten_anchors, flatten_bbox_preds, flatten_strides.unsqueeze(-1))flatten_preds[..., :4] = flatten_bboxesreturn flatten_predsmodel = YOLOV3().eval()
input = torch.zeros(1, 3, 416, 416, device='cpu')
torch.onnx.export(model, input, "out.onnx", opset_version=11)

导出的onnx结构如下:
在这里插入图片描述
安装mmdeploy的话,可以通过下面脚本导出onnx模型。

from mmdeploy.apis import torch2onnx
from mmdeploy.backend.sdk.export_info import export2SDKimg = 'bus.jpg'
work_dir = './work_dir/onnx/yolov3'
save_file = './end2end.onnx'
deploy_cfg = 'mmdeploy/configs/mmdet/detection/detection_onnxruntime_dynamic.py'
model_cfg = 'mmdetection/configs/yolo/yolov3_mobilenetv2_8xb24-ms-416-300e_coco.py'
model_checkpoint = 'checkpoints/yolov3_mobilenetv2_mstrain-416_300e_coco_20210718_010823-f68a07b3.pth'
device = 'cpu'# 1. convert model to onnx
torch2onnx(img, work_dir, save_file, deploy_cfg, model_cfg, model_checkpoint, device)# 2. extract pipeline info for sdk use (dump-info)
export2SDK(deploy_cfg, model_cfg, work_dir, pth=model_checkpoint, device=device)

onnx模型的结构如下:
在这里插入图片描述

onnxruntime推理

手动导出的onnx模型使用onnxruntime推理:

import cv2
import numpy as np
import onnxruntimeCLASSES = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light','fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow','elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee','skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard','tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple','sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch','potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone','microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear','hair drier', 'toothbrush'] #coco80类别       
use_letterbox = True
input_shape = (416, 416)      def nms(dets, thresh):x1 = dets[:, 0]y1 = dets[:, 1]x2 = dets[:, 2]y2 = dets[:, 3]areas = (y2 - y1 + 1) * (x2 - x1 + 1)scores = dets[:, 4]keep = []index = scores.argsort()[::-1] while index.size > 0:i = index[0]keep.append(i)x11 = np.maximum(x1[i], x1[index[1:]]) y11 = np.maximum(y1[i], y1[index[1:]])x22 = np.minimum(x2[i], x2[index[1:]])y22 = np.minimum(y2[i], y2[index[1:]])w = np.maximum(0, x22 - x11 + 1)                              h = np.maximum(0, y22 - y11 + 1) overlaps = w * hious = overlaps / (areas[i] + areas[index[1:]] - overlaps)idx = np.where(ious <= thresh)[0]index = index[idx + 1]return keepdef filter_box(org_box, conf_thres, iou_thres): #过滤掉无用的框org_box = np.squeeze(org_box)conf = org_box[..., 4] > conf_thresbox = org_box[conf == True]cls_cinf = box[..., 5:]cls = []for i in range(len(cls_cinf)):cls.append(int(np.argmax(cls_cinf[i])))all_cls = list(set(cls))    output = []for i in range(len(all_cls)):curr_cls = all_cls[i]curr_cls_box = []curr_out_box = []for j in range(len(cls)):if cls[j] == curr_cls:box[j][4] *= cls_cinf[j][curr_cls]box[j][5] = curr_clscurr_cls_box.append(box[j][:6])curr_cls_box = np.array(curr_cls_box)curr_out_box = nms(curr_cls_box,iou_thres)for k in curr_out_box:output.append(curr_cls_box[k])output = np.array(output) #(4, 6)return outputdef letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):# Resize and pad image while meeting stride-multiple constraintsshape = im.shape[:2]  # current shape [height, width]# Scale ratio (new / old)r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])# Compute paddingnew_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))    dw, dh = (new_shape[1] - new_unpad[0])/2, (new_shape[0] - new_unpad[1])/2  # wh padding top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))left, right = int(round(dw - 0.1)), int(round(dw + 0.1))if shape[::-1] != new_unpad:  # resizeim = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add borderreturn imdef scale_boxes(input_shape, boxes, shape):# Rescale boxes (xyxy) from input_shape to shapegain = min(input_shape[0] / shape[0], input_shape[1] / shape[1])  # gain  = old / newpad = (input_shape[1] - shape[1] * gain) / 2, (input_shape[0] - shape[0] * gain) / 2  # wh paddingboxes[..., [0, 2]] -= pad[0]  # x paddingboxes[..., [1, 3]] -= pad[1]  # y paddingboxes[..., :4] /= gainboxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])  # x1, x2boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])  # y1, y2return boxesdef draw(image, box_data):box_data = scale_boxes(input_shape, box_data, image.shape)boxes = box_data[...,:4].astype(np.int32) scores = box_data[...,4]classes = box_data[...,5].astype(np.int32)for box, score, cl in zip(boxes, scores, classes):top, left, right, bottom = boxprint('class: {}, score: {}, coordinate: [{}, {}, {}, {}]'.format(CLASSES[cl], score, top, left, right, bottom))cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 1)cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score), (top, left), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 1)if __name__=="__main__":img = cv2.imread('bus.jpg')if use_letterbox:input = letterbox(img, input_shape)else:input = cv2.resize(img, input_shape)input = input[:, :, ::-1].transpose(2, 0, 1).astype(dtype=np.float32)  #BGR2RGB和HWC2CHWinput[0,:] = (input[0,:] - 123.675) / 58.395   input[1,:] = (input[1,:] - 116.28) / 57.12input[2,:] = (input[2,:] - 103.53) / 57.375input = np.expand_dims(input, axis=0)onnx_session = onnxruntime.InferenceSession('yolov3.onnx', providers=['CPUExecutionProvider'])input_name=[]for node in onnx_session.get_inputs():input_name.append(node.name)output_name=[]for node in onnx_session.get_outputs():output_name.append(node.name)input_feed={}for name in input_name:input_feed[name] = inputpred = onnx_session.run(None, input_feed)outbox = filter_box(pred, 0.5, 0.5)draw(img, outbox)cv2.imwrite('res.jpg', img)

mmdeploy导出的onnx模型使用onnxruntime推理:

import cv2
import numpy as np
import onnxruntimeCLASSES = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light','fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow','elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee','skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard','tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple','sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch','potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone','microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear','hair drier', 'toothbrush'] #coco80类别       
use_letterbox = True
input_shape = (416, 416)      def filter_box(org_box, conf_thres): #删除置信度小于conf_thres的BOXflag = org_box[0][..., 4] > conf_thresbox = org_box[0][flag == True] cls = org_box[1][flag == True].reshape(-1, 1) output = np.concatenate((box, cls), axis=1)  return outputdef letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):# Resize and pad image while meeting stride-multiple constraintsshape = im.shape[:2]  # current shape [height, width]# Scale ratio (new / old)r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])# Compute paddingnew_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))    dw, dh = (new_shape[1] - new_unpad[0])/2, (new_shape[0] - new_unpad[1])/2  # wh padding top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))left, right = int(round(dw - 0.1)), int(round(dw + 0.1))if shape[::-1] != new_unpad:  # resizeim = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add borderreturn imdef scale_boxes(input_shape, boxes, shape):# Rescale boxes (xyxy) from input_shape to shapegain = min(input_shape[0] / shape[0], input_shape[1] / shape[1])  # gain  = old / newpad = (input_shape[1] - shape[1] * gain) / 2, (input_shape[0] - shape[0] * gain) / 2  # wh paddingboxes[..., [0, 2]] -= pad[0]  # x paddingboxes[..., [1, 3]] -= pad[1]  # y paddingboxes[..., :4] /= gainboxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])  # x1, x2boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])  # y1, y2return boxesdef draw(image, box_data):box_data = scale_boxes(input_shape, box_data, image.shape)boxes = box_data[...,:4].astype(np.int32) scores = box_data[...,4]classes = box_data[...,5].astype(np.int32)for box, score, cl in zip(boxes, scores, classes):top, left, right, bottom = boxprint('class: {}, score: {}, coordinate: [{}, {}, {}, {}]'.format(CLASSES[cl], score, top, left, right, bottom))cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 1)cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score), (top, left), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 1)if __name__=="__main__":img = cv2.imread('bus.jpg')if use_letterbox:input = letterbox(img, input_shape)else:input = cv2.resize(img, input_shape)input = input[:, :, ::-1].transpose(2, 0, 1).astype(dtype=np.float32)  #BGR2RGB和HWC2CHWinput[0,:] = (input[0,:] - 123.675) / 58.395   input[1,:] = (input[1,:] - 116.28) / 57.12input[2,:] = (input[2,:] - 103.53) / 57.375input = np.expand_dims(input, axis=0)onnx_session = onnxruntime.InferenceSession('../work_dir/onnx/yolov3/end2end.onnx', providers=['CPUExecutionProvider'])input_name=[]for node in onnx_session.get_inputs():input_name.append(node.name)output_name=[]for node in onnx_session.get_outputs():output_name.append(node.name)input_feed={}for name in input_name:input_feed[name] = inputpred = onnx_session.run(None, input_feed)outbox = filter_box(pred, 0.5)draw(img, outbox)cv2.imwrite('res.jpg', img)

直接使用mmdeploy的api推理:

from mmdeploy.apis import inference_modelmodel_cfg = 'mmdetection/configs/yolo/yolov3_mobilenetv2_8xb24-ms-416-300e_coco.py'
deploy_cfg = 'mmdeploy/configs/mmdet/detection/detection_onnxruntime_dynamic.py'
img = 'bus.jpg'
backend_files = ['work_dir/onnx/yolov3/end2end.onnx']
device = 'cpu'result = inference_model(model_cfg, deploy_cfg, backend_files, img, device)
print(result)

或者

from mmdeploy_runtime import Detector
import cv2# 读取图片
img = cv2.imread('bus.jpg')# 创建检测器
detector = Detector(model_path='work_dir/onnx/yolov3', device_name='cpu')# 执行推理
bboxes, labels, _ = detector(img)
# 使用阈值过滤推理结果,并绘制到原图中
indices = [i for i in range(len(bboxes))]
for index, bbox, label_id in zip(indices, bboxes, labels):[left, top, right, bottom], score = bbox[0:4].astype(int),  bbox[4]if score < 0.3:continuecv2.rectangle(img, (left, top), (right, bottom), (0, 255, 0))
cv2.imwrite('output_detection.png', img)

导出engine文件

这里通过trtexec转换onnx文件,LZ的版本是TensorRT-8.2.1.8。

./trtexec.exe --onnx=yolov3.onnx --saveEngine=yolov3.engine --workspace=20480

tensorrt推理

手动导出的模型使用tensorrt推理:

import cv2
import numpy as np
import tensorrt as trt
import pycuda.autoinit 
import pycuda.driver as cuda  CLASSES = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light','fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow','elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee','skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard','tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple','sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch','potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone','microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear','hair drier', 'toothbrush'] #coco80类别       
use_letterbox = True
input_shape = (416, 416)      def nms(dets, thresh):x1 = dets[:, 0]y1 = dets[:, 1]x2 = dets[:, 2]y2 = dets[:, 3]areas = (y2 - y1 + 1) * (x2 - x1 + 1)scores = dets[:, 4]keep = []index = scores.argsort()[::-1] while index.size > 0:i = index[0]keep.append(i)x11 = np.maximum(x1[i], x1[index[1:]]) y11 = np.maximum(y1[i], y1[index[1:]])x22 = np.minimum(x2[i], x2[index[1:]])y22 = np.minimum(y2[i], y2[index[1:]])w = np.maximum(0, x22 - x11 + 1)                              h = np.maximum(0, y22 - y11 + 1) overlaps = w * hious = overlaps / (areas[i] + areas[index[1:]] - overlaps)idx = np.where(ious <= thresh)[0]index = index[idx + 1]return keepdef filter_box(org_box, conf_thres, iou_thres): #过滤掉无用的框org_box = np.squeeze(org_box)conf = org_box[..., 4] > conf_thresbox = org_box[conf == True]cls_cinf = box[..., 5:]cls = []for i in range(len(cls_cinf)):cls.append(int(np.argmax(cls_cinf[i])))all_cls = list(set(cls))    output = []for i in range(len(all_cls)):curr_cls = all_cls[i]curr_cls_box = []curr_out_box = []for j in range(len(cls)):if cls[j] == curr_cls:box[j][4] *= cls_cinf[j][curr_cls]box[j][5] = curr_clscurr_cls_box.append(box[j][:6])curr_cls_box = np.array(curr_cls_box)curr_out_box = nms(curr_cls_box,iou_thres)for k in curr_out_box:output.append(curr_cls_box[k])output = np.array(output) #(4, 6)return outputdef letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):# Resize and pad image while meeting stride-multiple constraintsshape = im.shape[:2]  # current shape [height, width]# Scale ratio (new / old)r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])# Compute paddingnew_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))    dw, dh = (new_shape[1] - new_unpad[0])/2, (new_shape[0] - new_unpad[1])/2  # wh padding top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))left, right = int(round(dw - 0.1)), int(round(dw + 0.1))if shape[::-1] != new_unpad:  # resizeim = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add borderreturn imdef scale_boxes(input_shape, boxes, shape):# Rescale boxes (xyxy) from input_shape to shapegain = min(input_shape[0] / shape[0], input_shape[1] / shape[1])  # gain  = old / newpad = (input_shape[1] - shape[1] * gain) / 2, (input_shape[0] - shape[0] * gain) / 2  # wh paddingboxes[..., [0, 2]] -= pad[0]  # x paddingboxes[..., [1, 3]] -= pad[1]  # y paddingboxes[..., :4] /= gainboxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])  # x1, x2boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])  # y1, y2return boxesdef draw(image, box_data):box_data = scale_boxes(input_shape, box_data, image.shape)boxes = box_data[...,:4].astype(np.int32) scores = box_data[...,4]classes = box_data[...,5].astype(np.int32)for box, score, cl in zip(boxes, scores, classes):top, left, right, bottom = boxprint('class: {}, score: {}, coordinate: [{}, {}, {}, {}]'.format(CLASSES[cl], score, top, left, right, bottom))cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 1)cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score), (top, left), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 1)if __name__=="__main__":logger = trt.Logger(trt.Logger.WARNING)with open("yolov3.engine", "rb") as f, trt.Runtime(logger) as runtime:engine = runtime.deserialize_cuda_engine(f.read())context = engine.create_execution_context()h_input = cuda.pagelocked_empty(trt.volume(context.get_binding_shape(0)), dtype=np.float32)h_output = cuda.pagelocked_empty(trt.volume(context.get_binding_shape(1)), dtype=np.float32)d_input = cuda.mem_alloc(h_input.nbytes)d_output = cuda.mem_alloc(h_output.nbytes)stream = cuda.Stream()img = cv2.imread('bus.jpg')if use_letterbox:input = letterbox(img, input_shape)else:input = cv2.resize(img, input_shape)input = input[:, :, ::-1].transpose(2, 0, 1).astype(dtype=np.float32)  #BGR2RGB和HWC2CHWinput[0,:] = (input[0,:] - 123.675) / 58.395   input[1,:] = (input[1,:] - 116.28) / 57.12input[2,:] = (input[2,:] - 103.53) / 57.375input = np.expand_dims(input, axis=0) np.copyto(h_input, input.ravel())with engine.create_execution_context() as context:cuda.memcpy_htod_async(d_input, h_input, stream)context.execute_async_v2(bindings=[int(d_input), int(d_output)], stream_handle=stream.handle)cuda.memcpy_dtoh_async(h_output, d_output, stream)stream.synchronize()  pred = h_output.reshape(1, 10647, 85) outbox = filter_box(pred, 0.5, 0.5)draw(img, outbox)cv2.imwrite('res.jpg', img)

使用mmdeploy的api推理:

from mmdeploy.apis import inference_modelmodel_cfg ='mmdetection/configs/yolo/yolov3_mobilenetv2_8xb24-ms-416-300e_coco.py'
deploy_cfg = 'mmdeploy/configs/mmdet/detection/detection_tensorrt_static-320x320.py'
img = 'bus.jpg'
backend_files = ['work_dir/trt/yolov3/end2end.engine']
device = 'cuda'result = inference_model(model_cfg, deploy_cfg, backend_files, img, device)
print(result)

或者

from mmdeploy_runtime import Detector
import cv2# 读取图片
img = cv2.imread('bus.jpg')# 创建检测器
detector = Detector(model_path='work_dir/trt/yolox', device_name='cuda')# 执行推理
bboxes, labels, _ = detector(img)
# 使用阈值过滤推理结果,并绘制到原图中
indices = [i for i in range(len(bboxes))]
for index, bbox, label_id in zip(indices, bboxes, labels):[left, top, right, bottom], score = bbox[0:4].astype(int),  bbox[4]if score < 0.3:continuecv2.rectangle(img, (left, top), (right, bottom), (0, 255, 0))
cv2.imwrite('output_detection.png', img)

相关文章:

OpenMMlab导出yolov3模型并用onnxruntime和tensorrt推理

导出onnx文件 直接使用脚本 import torch from mmdet.apis import init_detector, inference_detectorconfig_file ./configs/yolo/yolov3_mobilenetv2_8xb24-ms-416-300e_coco.py checkpoint_file yolov3_mobilenetv2_mstrain-416_300e_coco_20210718_010823-f68a07b3.pth…...

单链表的插入删除

#include <iostream>#include <stdio.h> #include <stdlib.h>using namespace std;//带头指针的单链表typedef struct LNode{int data;struct LNode *next;}LNode, *LinkList;bool InitList(LinkList &L){L (LNode *) malloc(sizeof(LNode));if(L NUL…...

github使用手册

核心代码 配置用户名/邮箱 best practice git init #在本地初始化一个仓库 git add . #将当前目录所有的文件加入&#xff08;注意这里是加入&#xff09;到缓存区 git commit -m "xxx" #将当前缓存区里的内容提交到本地仓库 git remote add <remote_rep_name&g…...

怎样做ChatGPT应用开发?

要开发一个基于ChatGPT的应用&#xff0c;你可以按照以下步骤进行&#xff1a; 步骤1&#xff1a;了解ChatGPT API ChatGPT的使用通常通过API进行。你需要了解ChatGPT的API文档&#xff0c;包括如何进行请求、API端点、身份验证等信息。在开发之前&#xff0c;确保你已经获取了…...

漏洞-任意账号注册

一漏洞介绍 1.未验证邮箱/手机号 情景&#xff1a;应用为了方便用户记录用户名&#xff0c;使用邮箱和手机号作为用户名&#xff08;因此很多应用在注册的时候就要求用户填写&#xff0c;多数时候都会给用户发送激活信息&#xff0c;激活后才能登录&#xff09; 缺陷&#xff…...

一个关于jdbc操作mysql和java基础练手的通讯录管理系统小项目

首先 : 整个项目的项目结构为 : 1.第一步先导入数据库的驱动&#xff0c;我的mysql数据库是8.0以上版本&#xff0c;然后导入的驱动就是8.0.16版本的jar包&#xff1b; 1.JdbcBase : JDBC基础操作封装成了JdbcBase类,在里面先静态定义了数据库连接对象和DQL查询结果&#x…...

C++用条件变量实现线程安全的queue容器

#include <queue> #include <memory> #include <mutex> #include <condition_variable> template<typename T> class threadsafe_queue { private:mutable std::mutex mut; // 1 互斥量必须是可变的 std::queue<T> data_queue;std::condi…...

EDA实验-----3-8译码器设计(QuartusII)

目录 一. 实验目的 二. 实验仪器 三. 实验原理及内容 1.实验原理 2.实验内容 四&#xff0e;实验步骤 五. 实验报告 六. 注意事项 七. 实验过程 1.创建Verilog文件&#xff0c;写代码 ​编辑 2.波形仿真 3.连接电路图 4.烧录操作 一. 实验目的 学会Verilog HDL的…...

NFTScan | 11.06~11.12 NFT 市场热点汇总

欢迎来到由 NFT 基础设施 NFTScan 出品的 NFT 生态热点事件每周汇总。 周期&#xff1a;2023.11.06~ 2023.11.12 NFT Hot News 01/ 《辛普森一家》提及 NFT 及区块链&#xff0c;相关 NFT 地板价涨至 0.35 ETH 11 月 6 日&#xff0c;据 Coindesk 报道&#xff0c;美国时间周…...

2022年12月 Python(五级)真题解析#中国电子学会#全国青少年软件编程等级考试

Python等级考试(1~6级)全部真题・点这里 一、单选题(共25题,每题2分,共50分) 第1题 下面哪个语句正确定义了元组类型数据tuple1?( ) A: tuple1=[“张三”,“李四”,“王五”] B: tuple1=(“张三”;“李四”;“王五”) C: tuple1=(张三,李四,王五) D: tuple1=(“张三…...

第三章 将对象映射到 XML - 使用列表或数组定义的属性

文章目录 第三章 将对象映射到 XML - 使用列表或数组定义的属性使用列表或数组定义的属性%ListOfDataTypes 或 %ArrayOfDataTypes 类型的属性%ListOfObjects 或 %ArrayOfObjects 类型的属性 第三章 将对象映射到 XML - 使用列表或数组定义的属性 使用列表或数组定义的属性 对…...

C/S架构学习之基于TCP的本地通信(客户机)

基于TCP的本地通信&#xff08;客户机&#xff09;&#xff1a;创建流程&#xff1a;一、创建字节流式套接字&#xff08;socket函数&#xff09;&#xff1a; int sock_fd socket(AF_LOCAL,SOCK_STREAM,0);二、创建客户机和服务器的本地网络信息结构体并填充客户机和服务器本…...

CCF 备忘

一、不错的网站 CCF CCSP 竞赛历年资料 官网 http://118.190.20.162/home.page 二、基础套路 循环输入 数组标记法&#xff08;数组下标-数值 的映射&#xff09; 两个矩阵相乘 map<long long, map<long long, long long> > ans; for(int i1;i<d;i){for(int…...

Spring Framework中的依赖注入:构造器注入 vs. Setter注入

前言 构造器注入和Setter注入是依赖注入&#xff08;Dependency Injection&#xff0c;DI&#xff09;中两种常见的方式&#xff0c;用于向一个对象注入其所依赖的其他对象或数值。这两种注入方式有各自的特点和用途。 构造器注入&#xff08;Constructor Injection&#xff…...

Java学习之路 —— API篇

文章目录 前言Object类2. Objects类3. 包装类4. StringBuilder和StringBuffer5. StringJoiner6. Math7. System8. JDK8开始新增的日期、时间9. Arrays10. Lambda表达式11. 方法引用 前言 其实转语言来说&#xff0c;语法都比较简单&#xff0c;花个三天就会了&#xff0c;但最…...

Windows下安装Anaconda5.3.1+Python3.8+TensorFlow2.13.0-CPU版本总结

Python3.8安装可以参考博文https://janus.blog.csdn.net/article/details/55274849 进行安装即可。 【1】Anaconda 清华的开源软件镜像站&#xff1a;https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive/下载&#xff0c;这里选择的是5.3.1版本。 然后正常安装就可以&am…...

DigitalVirt 日本 Lite VPS 测评

发布于 2023-07-18 在 https://chenhaotian.top/vps/digitalvirt-jp-lite/ 官网链接&#xff08;含AFF&#xff09;&#xff1a;https://digitalvirt.com/aff.php?aff459 日本 软银 CMI&#xff0c;联通首选。 流媒体解锁很好&#xff0c;详见测试。 注意&#xff0c;测速…...

Ridgeline plot / 远山图 / 山脊图 怎么画?怎么优化?

工具 Origin 2022 当然&#xff0c;用Matlab、Python也是可以的。 颜色配置 色卡调整...

【STM32/FreeRTOS】SysTick定时器及FreeRTOS系统节拍

目录 一、SysTick定时器 1、SysTick寄存器介绍 &#xff08;1&#xff09;控制及状态寄存器 &#xff08;2&#xff09;重装载数值寄存器 &#xff08;3&#xff09;当前数值寄存器 2、SysTick寄存器配置函数 二、FreeRTOS中的SysTick定时器 1、SysTick配置函数及分析 …...

Vue3封装自定义指令+h()

官方install介绍 directive/myDir/index.js 定义指令 import { h, render, ref } from "vue"; const vMyDir {mounted(el, binding) {renderElement(el, binding);}, }; // inserted是Vue2的生命周期钩子&#xff0c;所以在Vue3项目中要使用mounted const renderEl…...

KubeSphere 容器平台高可用:环境搭建与可视化操作指南

Linux_k8s篇 欢迎来到Linux的世界&#xff0c;看笔记好好学多敲多打&#xff0c;每个人都是大神&#xff01; 题目&#xff1a;KubeSphere 容器平台高可用&#xff1a;环境搭建与可视化操作指南 版本号: 1.0,0 作者: 老王要学习 日期: 2025.06.05 适用环境: Ubuntu22 文档说…...

Unity3D中Gfx.WaitForPresent优化方案

前言 在Unity中&#xff0c;Gfx.WaitForPresent占用CPU过高通常表示主线程在等待GPU完成渲染&#xff08;即CPU被阻塞&#xff09;&#xff0c;这表明存在GPU瓶颈或垂直同步/帧率设置问题。以下是系统的优化方案&#xff1a; 对惹&#xff0c;这里有一个游戏开发交流小组&…...

定时器任务——若依源码分析

分析util包下面的工具类schedule utils&#xff1a; ScheduleUtils 是若依中用于与 Quartz 框架交互的工具类&#xff0c;封装了定时任务的 创建、更新、暂停、删除等核心逻辑。 createScheduleJob createScheduleJob 用于将任务注册到 Quartz&#xff0c;先构建任务的 JobD…...

12.找到字符串中所有字母异位词

&#x1f9e0; 题目解析 题目描述&#xff1a; 给定两个字符串 s 和 p&#xff0c;找出 s 中所有 p 的字母异位词的起始索引。 返回的答案以数组形式表示。 字母异位词定义&#xff1a; 若两个字符串包含的字符种类和出现次数完全相同&#xff0c;顺序无所谓&#xff0c;则互为…...

Java求职者面试指南:计算机基础与源码原理深度解析

Java求职者面试指南&#xff1a;计算机基础与源码原理深度解析 第一轮提问&#xff1a;基础概念问题 1. 请解释什么是进程和线程的区别&#xff1f; 面试官&#xff1a;进程是程序的一次执行过程&#xff0c;是系统进行资源分配和调度的基本单位&#xff1b;而线程是进程中的…...

【Linux】Linux 系统默认的目录及作用说明

博主介绍&#xff1a;✌全网粉丝23W&#xff0c;CSDN博客专家、Java领域优质创作者&#xff0c;掘金/华为云/阿里云/InfoQ等平台优质作者、专注于Java技术领域✌ 技术范围&#xff1a;SpringBoot、SpringCloud、Vue、SSM、HTML、Nodejs、Python、MySQL、PostgreSQL、大数据、物…...

【 java 虚拟机知识 第一篇 】

目录 1.内存模型 1.1.JVM内存模型的介绍 1.2.堆和栈的区别 1.3.栈的存储细节 1.4.堆的部分 1.5.程序计数器的作用 1.6.方法区的内容 1.7.字符串池 1.8.引用类型 1.9.内存泄漏与内存溢出 1.10.会出现内存溢出的结构 1.内存模型 1.1.JVM内存模型的介绍 内存模型主要分…...

PostgreSQL——环境搭建

一、Linux # 安装 PostgreSQL 15 仓库 sudo dnf install -y https://download.postgresql.org/pub/repos/yum/reporpms/EL-$(rpm -E %{rhel})-x86_64/pgdg-redhat-repo-latest.noarch.rpm# 安装之前先确认是否已经存在PostgreSQL rpm -qa | grep postgres# 如果存在&#xff0…...

在Zenodo下载文件 用到googlecolab googledrive

方法&#xff1a;Figshare/Zenodo上的数据/文件下载不下来&#xff1f;尝试利用Google Colab &#xff1a;https://zhuanlan.zhihu.com/p/1898503078782674027 参考&#xff1a; 通过Colab&谷歌云下载Figshare数据&#xff0c;超级实用&#xff01;&#xff01;&#xff0…...

Java中HashMap底层原理深度解析:从数据结构到红黑树优化

一、HashMap概述与核心特性 HashMap作为Java集合框架中最常用的数据结构之一&#xff0c;是基于哈希表的Map接口非同步实现。它允许使用null键和null值&#xff08;但只能有一个null键&#xff09;&#xff0c;并且不保证映射顺序的恒久不变。与Hashtable相比&#xff0c;Hash…...