yolov5 dpu

### 如何在DPU上运行YOLOv5 为了使YOLOv5能够在DPU（深度学习处理单元）上高效运行，需完成一系列特定配置和优化工作。这不仅涉及模型转换还涉及到环境搭建以及性能调优。 #### 环境准备 安装必要的依赖库对于确保YOLOv5能在DPU环境中顺利执行至关重要。通常情况下，需要先设置好支持Vitis AI开发套件的Python虚拟环境，并安装PyTorch框架以及其他辅助工具[^1]。 ```bash conda create -n vitis-ai python=3.7 conda activate vitis-ai pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu ``` #### 模型转换 由于原始版本的YOLOv5并非专门为DPU设计，在将其部署到此类硬件之前可能需要通过量化、剪枝等方式来调整网络结构以便更好地适配目标平台特性并提高推理速度。具体来说，可以利用Xilinx提供的`vai_q_pytorch`命令行工具来进行模型压缩与编译操作[^2]。 ```python from pathlib import Path import sys sys.path.append(str(Path().cwd())) # 导入所需的模块 from models.experimental import attempt_load from utils.datasets import LoadImages, letterbox from utils.general import check_img_size, non_max_suppression, scale_coords from utils.torch_utils import select_device, time_synchronized weights = 'yolov5s.pt' # 权重文件路径 imgsz = 640 # 输入图片尺寸 device = select_device('') # 自动选择可用设备(CUDA优先) model = attempt_load(weights, map_location=device) # 加载预训练权重 if device.type != 'cpu': model.half() # 将FP32转为FP16半精度浮点数以加速计算过程 # 使用Vitis-AI工具链对模型进行量化感知训练(QAT)，从而获得更高效的INT8整数量化版YOLOv5 !vai_q_pytorch quantize \ --input_frozen_graph ./quantization_results/yolov5_quantized.pth \ --calib_iter 100 \ --output_dir ./quantization_results/ ``` #### 性能评估与调试 一旦完成了上述准备工作之后，则可以通过编写简单的测试脚本来验证最终效果如何。此时应该关注几个方面：一是确认检测结果准确性；二是测量实际耗时情况；三是观察资源占用率等指标表现是否满足预期要求[^3]。 ```python import cv2 import numpy as np from PIL import ImageDraw, ImageFont conf_thres = 0.4 # 置信度阈值 iou_thres = 0.5 # IOU交并比阈值用于NMS非极大抑制算法过滤重复框体 names = ['person', ... ] # 类别名称列表(可根据实际情况修改) def plot_one_box(x, img, color=None, label=None, line_thickness=None): tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3])) cv2.rectangle(img, c1, c2, color=color, thickness=tl) if label: tf = max(tl - 1, 1) t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3 cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA) cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA) source_image_path = "test.jpg" image = cv2.imread(source_image_path) height, width = image.shape[:2] # 对输入图像做适当缩放和平移变换使其适应于YOLOv5输入规格 scaled_image = letterbox(image, new_shape=imgsz)[0] scaled_image = scaled_image[:, :, ::-1].transpose(2, 0, 1).copy() scaled_image = np.ascontiguousarray(scaled_image.astype(np.float32)) scaled_image /= 255.0 with torch.no_grad(): pred = model(torch.from_numpy(scaled_image).to(device).unsqueeze_(dim=0))[0] detected_objects = [] for i, det in enumerate(non_max_suppression(pred, conf_thres, iou_thres)): gn = torch.tensor([width, height, width, height]).to(device) if len(det): det[:, :4] = scale_coords(scaled_image.shape[1:], det[:, :4], image.shape).round() for *xyxy, conf, cls in reversed(det): detected_objects.append((cls.item(), xyxy)) draw = ImageDraw.Draw(Image.open(source_image_path)) font = ImageFont.load_default() for obj_class_id, box in detected_objects: class_name = names[int(obj_class_id)] draw.text((box[0]+5, box[1]-10), f"{class_name} {conf:.2f}", fill=(255, 0, 0), font=font) draw.rectangle(box, outline="red", width=3) result_image.save("detection_result.png") ```