OpenMMlab导出yolov3模型并用onnxruntime和tensorrt推理

导出onnx文件

直接使用脚本

python">import torch
from mmdet.apis import init_detector, inference_detector


config_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：

python">import torch
from itertools import repeat
from mmdet.apis import init_detector, inference_detector


config_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 = 80
        self.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 = center
            base_anchors = []
            for base_size in base_sizes_per_level:
                w, h = base_size
                base_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_anchors
             
    def _meshgrid(self, x, y):
        xx = x.repeat(y.shape[0])
        yy = y.view(-1, 1).repeat(1, x.shape[0]).view(-1)
        return xx, yy

    def 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_w
            shift_y = torch.arange(0, feat_h) * stride_h
            shift_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_anchors
    
    def decode(self, bboxes, pred_bboxes, stride):
        xy_centers = (bboxes[..., :2] + bboxes[..., 2:]) * 0.5 + (pred_bboxes[..., :2] - 0.5) * stride
        whs = (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_bboxes
        
    def 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_objectness
        flatten_cls_scores = flatten_preds[..., 5:].sigmoid()
        flatten_preds[..., 5:] = flatten_cls_scores

        featmap_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_bboxes
        
        return flatten_preds
    
    
model = 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。

python">from mmdeploy.apis import torch2onnx
from mmdeploy.backend.sdk.export_info import export2SDK


img = '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：

python">import cv2
import numpy as np
import onnxruntime


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 * h
        ious = overlaps / (areas[i] + areas[index[1:]] - overlaps)
        idx = np.where(ious <= thresh)[0]
        index = index[idx + 1]
    return keep


def 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_thres
    box = 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_cls
                curr_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 output


def letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    
    # Compute padding
    new_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:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return im


def scale_boxes(input_shape, boxes, shape):
    # Rescale boxes (xyxy) from input_shape to shape
    gain = min(input_shape[0] / shape[0], input_shape[1] / shape[1])  # gain  = old / new
    pad = (input_shape[1] - shape[1] * gain) / 2, (input_shape[0] - shape[0] * gain) / 2  # wh padding

    boxes[..., [0, 2]] -= pad[0]  # x padding
    boxes[..., [1, 3]] -= pad[1]  # y padding
    boxes[..., :4] /= gain
    boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])  # x1, x2
    boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])  # y1, y2
    return boxes


def 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 = box
        print('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HWC2CHW
    input[0,:] = (input[0,:] - 123.675) / 58.395   
    input[1,:] = (input[1,:] - 116.28) / 57.12
    input[2,:] = (input[2,:] - 103.53) / 57.375
    input = 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] = input
        
    pred = 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：

python">import cv2
import numpy as np
import onnxruntime


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 filter_box(org_box, conf_thres): #删除置信度小于conf_thres的BOX
    flag = org_box[0][..., 4] > conf_thres
    box = org_box[0][flag == True] 
    cls = org_box[1][flag == True].reshape(-1, 1) 
    output = np.concatenate((box, cls), axis=1)  
    return output


def letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    
    # Compute padding
    new_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:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return im


def scale_boxes(input_shape, boxes, shape):
    # Rescale boxes (xyxy) from input_shape to shape
    gain = min(input_shape[0] / shape[0], input_shape[1] / shape[1])  # gain  = old / new
    pad = (input_shape[1] - shape[1] * gain) / 2, (input_shape[0] - shape[0] * gain) / 2  # wh padding

    boxes[..., [0, 2]] -= pad[0]  # x padding
    boxes[..., [1, 3]] -= pad[1]  # y padding
    boxes[..., :4] /= gain
    boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])  # x1, x2
    boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])  # y1, y2
    return boxes


def 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 = box
        print('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HWC2CHW
    input[0,:] = (input[0,:] - 123.675) / 58.395   
    input[1,:] = (input[1,:] - 116.28) / 57.12
    input[2,:] = (input[2,:] - 103.53) / 57.375
    input = 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] = input
        
    pred = 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：

python">from mmdeploy.apis import inference_model


model_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

python">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:
      continue
  cv2.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。

python">./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：

python">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 * h
        ious = overlaps / (areas[i] + areas[index[1:]] - overlaps)
        idx = np.where(ious <= thresh)[0]
        index = index[idx + 1]
    return keep


def 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_thres
    box = 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_cls
                curr_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 output


def letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    
    # Compute padding
    new_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:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return im


def scale_boxes(input_shape, boxes, shape):
    # Rescale boxes (xyxy) from input_shape to shape
    gain = min(input_shape[0] / shape[0], input_shape[1] / shape[1])  # gain  = old / new
    pad = (input_shape[1] - shape[1] * gain) / 2, (input_shape[0] - shape[0] * gain) / 2  # wh padding

    boxes[..., [0, 2]] -= pad[0]  # x padding
    boxes[..., [1, 3]] -= pad[1]  # y padding
    boxes[..., :4] /= gain
    boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])  # x1, x2
    boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])  # y1, y2
    return boxes


def 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 = box
        print('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HWC2CHW
    input[0,:] = (input[0,:] - 123.675) / 58.395   
    input[1,:] = (input[1,:] - 116.28) / 57.12
    input[2,:] = (input[2,:] - 103.53) / 57.375
    input = 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：

python">from mmdeploy.apis import inference_model


model_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

python">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:
      continue
  cv2.rectangle(img, (left, top), (right, bottom), (0, 255, 0))
cv2.imwrite('output_detection.png', img)