前言
最近在研究如何让YOLOv5推理得更快，总体看来，主要有以下这些思路：
（图片来源网络，侵删）
- 使用更快的 GPU，即：P100 -> V100 -> A100
- 多卡GPU推理
- 减小模型尺寸，即YOLOv5x -> YOLOv5l -> YOLOv5m -> YOLOv5s -> YOLOv5n
- 进行半精度FP16推理与python detect.py --half
- 减少–img-size，即 1280 -> 640 -> 320
- 导出成ONNX或OpenVINO格式，获得CPU加速
- 导出到TensorRT获得GPU加速
- 批量输入图片进行推理
- 使用多进程/多线程进行推理
注：使用多卡GPU和多进程/多线程的推理并不会对单张图片推理起到加速作用，只适用于很多张图片一起进行推理的场景。
本篇主要来研究多进程/多线程是否能对YOLOv5算法推理起到加速作用。
（图片来源网络，侵删）实验环境
GPU：RTX2060
torch：1.7.1+cu110
检测图片大小：1920x1080
img-size：1920
使用半精度推理half=True
推理模型：yolov5m.pt
实验过程
先放实验代码(detect.py)，根据官方源码进行了小改：
import configparser import time from pathlib import Path import cv2 import torch import threading import sys import multiprocessing as mp sys.path.append("yolov5") from models.experimental import attempt_load from utils.datasets import LoadImages from utils.general import check_img_size, non_max_suppression, scale_coords from utils.plots import Annotator, colors from utils.torch_utils import select_device from concurrent.futures import ThreadPoolExecutor Detect_path = 'D:/Data/detect_outputs' # 检测图片输出路径 def detect(path, model_path, detect_size): source = path weights = model_path imgsz = detect_size conf_thres = 0.25 iou_thres = 0.45 device = "" augment = True save_img = True save_dir = Path(Detect_path) # increment run device = select_device(device) half = device.type != 'cpu' # half precision only supported on CUDA # Load model model = attempt_load(weights, map_location=device) # load FP32 model stride = int(model.stride.max()) # model stride imgsz = check_img_size(imgsz, s=stride) # check img_sizef if half: model.half() # to FP16 # Set Dataloader vid_path, vid_writer = None, None dataset = LoadImages(source, img_size=imgsz, stride=stride) # Get names and colors names = model.module.names if hasattr(model, 'module') else model.names # Run inference if device.type != 'cpu': model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters()))) # run once result_list = [] for path, img, im0s, vid_cap in dataset: # 读取图片传到gpu上 t1 = time.time() img = torch.from_numpy(img).to(device) print("read pictures cost time：", time.time() - t1) t2 = time.time() img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 if img.ndimension() == 3: img = img.unsqueeze(0) print("process pictures cost time：", time.time() - t2) # Inference pred = model(img, augment=augment)[0] # Apply NMS pred = non_max_suppression(pred, conf_thres, iou_thres) # Process detections for i, det in enumerate(pred): # detections per image p, s, im0, frame = path, '', im0s, getattr(dataset, 'frame', 0) p = Path(p) # to Path save_path = str(save_dir / p.name) # img.jpg s += '%gx%g ' % img.shape[2:] # print string # print(s) # 384x640 s_result = '' # 输出检测结果 annotator = Annotator(im0, line_width=3, example=str(names)) if len(det): # Rescale boxes from img_size to im0 size det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round() # Print results for c in det[:, -1].unique(): n = (det[:, -1] == c).sum() # detections per class # s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string s += f"{n} {names[int(c)]}, " # add to string s_result += f"{n} {names[int(c)]} " # Write results for *xyxy, conf, cls in reversed(det): if save_img: c = int(cls) # label = f'{names[int(cls)]} {conf:.2f}' label = f'{names[int(cls)]}' # print(label) annotator.box_label(xyxy, label, color=colors(c, True)) # print(xyxy) print(f'{s}') # print(f'{s_result}') result_list.append(s_result) # 将conf对象中的数据写入到文件中 conf = configparser.ConfigParser() cfg_file = open("glovar.cfg", 'w') conf.add_section("default") # 在配置文件中增加一个段 # 第一个参数是段名，第二个参数是选项名，第三个参数是选项对应的值 conf.set("default", "process", str(dataset.img_count)) conf.set("default", "total", str(dataset.nf)) conf.write(cfg_file) cfg_file.close() im0 = annotator.result() # Save results (image with detections) t3 = time.time() if save_img: if dataset.mode == 'image': cv2.imwrite(save_path, im0) else: # 'video' or 'stream' if vid_path != save_path: # new video vid_path = save_path if isinstance(vid_writer, cv2.VideoWriter): vid_writer.release() # release previous video writer if vid_cap: # video fps = vid_cap.get(cv2.CAP_PROP_FPS) w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) else: # stream fps, w, h = 30, im0.shape[1], im0.shape[0] save_path += '.mp4' vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) vid_writer.write(im0) print("write pictures cost time：", time.time() - t3) print('Done') def run(path, model_path, detect_size): with torch.no_grad(): detect(path, model_path, detect_size)
首先进行小批量的图片进行实验，下面输入两张图片进行检测。
原始推理
if __name__ == '__main__': s_t = time.time() path1 = "D:/Data/image/DJI_0001_00100.jpg" path2 = "D:/Data/image/DJI_0001_00530.jpg" model_path = "../weights/best.pt" detect_size = 1920 run(path1, model_path, detect_size) run(path2, model_path, detect_size) print("Tatal Cost Time:", time.time() - s_t)Tatal Cost Time: 3.496427059173584
线程池推理
开辟两个线程进行推理：
if __name__ == '__main__': s_t = time.time() pool = ThreadPoolExecutor(max_workers=2) path1 = "D:/Data/image/DJI_0001_00100.jpg" path2 = "D:/Data/image/DJI_0001_00530.jpg" model_path = "../weights/best.pt" detect_size = 1920 pool.submit(run, path1, model_path, detect_size) pool.submit(run, path2, model_path, detect_size) pool.shutdown(wait=True) print("Tatal Cost Time:", time.time() - s_t)Tatal Cost Time: 3.2433135509490967
开双线程推理和原始推理时间类似，再次验证了python中的”伪多线程”。
进程池推理
开辟两个进程进行推理：
if __name__ == '__main__': s_t = time.time() pool = mp.Pool(processes=2) path1 = "D:/Data/image/DJI_0001_00100.jpg" path2 = "D:/Data/image/DJI_0001_00530.jpg" model_path = "../weights/best.pt" detect_size = 1920 pool.apply_async(run, (path1, model_path, detect_size,)) pool.apply_async(run, (path2, model_path, detect_size,)) pool.close() pool.join() print("Tatal Cost Time:", time.time() - s_t)Tatal Cost Time: 6.020772695541382
双进程推理
双进程推理时间竟然是原始推理的两倍，以为是进程池的开销太大，于是换种写法，不使用进程池：
if __name__ == '__main__': s_t = time.time() path1 = "D:/Data/image/DJI_0001_00100.jpg" path2 = "D:/Data/image/DJI_0001_00530.jpg" model_path = "../weights/best.pt" detect_size = 1920 p1 = mp.Process(target=run, args=(path1, model_path, detect_size,)) p2 = mp.Process(target=run, args=(path2, model_path, detect_size,)) p1.start() p2.start() p1.join() p2.join() print("Tatal Cost Time:", time.time() - s_t)Tatal Cost Time: 6.089479446411133
发现双进程时间仍然较久，说明在数据较少时，进程的开销成本过高，这和我之前做的实验多线程和多进程的效率对比结果相类似。
于是下面将图像数量扩大到300张进行实验。
300pic-原始推理
if __name__ == '__main__': s_t = time.time() path1 = "D:/Data/image" path2 = "D:/Data/image2" path3 = "D:/Data/image3" model_path = "../weights/best.pt" detect_size = 1920 run(path1, model_path, detect_size) run(path2, model_path, detect_size) run(path3, model_path, detect_size) print("Tatal Cost Time:", time.time() - s_t)Tatal Cost Time: 62.02898120880127
300pic-多进程推理
if __name__ == '__main__': s_t = time.time() path1 = "D:/Data/image" path2 = "D:/Data/image2" path3 = "D:/Data/image3" model_path = "../weights/best.pt" detect_size = 1920 p1 = mp.Process(target=run, args=(path1, model_path, detect_size,)) p2 = mp.Process(target=run, args=(path2, model_path, detect_size,)) p3 = mp.Process(target=run, args=(path3, model_path, detect_size,)) p1.start() p2.start() p3.start() p1.join() p2.join() p3.join() print("Tatal Cost Time:", time.time() - s_t)Tatal Cost Time: 47.85872721672058
和预期一样，当数据量提升上去时，多进程推理的速度逐渐超越原始推理。
总结
本次实验结果如下表所示：
图像处理张数 原始推理(s) 多线程推理(s) 多进程推理(s) 2 3.49 3.24 6.08 300 62.02 / 47.85 值得注意的是，使用多进程推理时，进程间保持独立，这意味着模型需要被重复在GPU上进行创建，因此，可以根据单进程所占显存大小来估算显卡所支持的最大进程数。
后续：在顶配机上进行实验
后面嫖到了组里i9-13700K+RTX4090的顶配主机，再进行实验，结果如下：
图像处理张数 原始推理(s) 多线程推理(s) 多进程推理(s) 2 2.21 2.09 3.92 300 29.23 / 17.61 后记：更正结论
后面觉得之前做的实验有些草率，尽管Python存在GIL的限制，但是在此类IO频繁的场景中，多线程仍然能缓解IO阻塞，从而实现加速，因此选用YOLOv5s模型，在4090上，对不同分辨率的图片进行测试：
输入图像分辨率：1920x1080
图像数量 原始推理(s) 双线程推理(s) 双进程推理(s) 2 1.92 1.85 3.92 100 7.02 4.91 6.52 200 13.07 8.10 9.66 输入图像分辨率：13400x9528
图像数量 原始推理(s) 双线程推理(s) 双进程推理(s) 2 6.46 4.99 7.03 100 190.85 119.43 117.12 200 410.95 239.84 239.51
