撰文｜FengWen、BBuf

代码仓库地址：
github.com/Oneflow-Inc… 欢迎star one-yolov5项目获取最新的动态。 如果你有问题，欢迎在仓库给我们提出宝贵的意见。如果对你有帮助，欢迎来给我Star呀~

1

模型融合 (Model Ensembling)

融合建模（Ensemble modeling）是这样一个过程：通过使用许多不同的建模算法或使用不同的训练数据集，创建多个不同的模型来预测结果。然后，融合模型汇总每个基础模型（base model）的预测结果，并对未见过的数据进行最终预测。使用融合模型的动机是为了减少预测的泛化误差。只要基础模型具有多样性和独立性，采用融合方法时，模型的预测误差就会减小。这种方法在做出预测时寻求集体智慧（wisdom of crowd）。尽管融合模型在模型中具有多个基础模型，但它就像单个模型那样运行和执行。（
www.sciencedirect.com/topics/comp…

这个教程用来解释在YOLOv5模型的测试和推理中如何使用模型融合 (Model Ensembling)提高mAP和Recall。欢迎大家到这里查看本篇文章的完整版本：start.oneflow.org/oneflow-yol…

2

开始之前

克隆工程并在 Python>3.7.0 的环境中安装 requiresments.txt , OneFlow 请选择 nightly 版本或者 >0.9 版本 。模型和数据可以从源码中自动下载。

git clone https://github.com/Oneflow-Inc/one-yolov5.git
cd one-yolov5
pip install -r requirements.txt  # install

3

普通测试

在尝试TTA之前，我们希望建立一个基准能够进行比较。该命令在COCO val2017上以640像素的图像大小测试YOLOv5x。yolov5x 是可用的最大并且最精确的模型。其它可用的模型是 yolov5s, yolov5m 和 yolov5l等 或者 自己从数据集训练出的模型 ./weights/best。有关所有可用模型的详细信息，请参阅我们的 READEME table

$ python val.py --weights ./yolov5x --data coco.yaml --img 640 --half

输出:

val: data=data/coco.yaml, weights=['./yolov5x'], batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, task=val, device=, workers=8, single_cls=False, augment=False, verbose=False, save_txt=False, save_hybrid=False, save_conf=False, save_json=True, project=runs/val, name=exp, exist_ok=False, half=True, dnn=False
YOLOv5  v1.0-8-g94ec5c4 Python-3.8.13 oneflow-0.8.1.dev20221018+cu112 
Fusing layers... 
Model summary: 322 layers, 86705005 parameters, 571965 gradients
val: Scanning 'https://b2.7b2.com/data/dataset/fengwen/coco/val2017.cache' images and labels... 4952 found, 48 missing, 0 empty, 0 corrupt: 100%|████████
               Class     Images     Labels          P          R     mAP@.5 mAP@.5:.95: 100%|██████████| 157/157 [01:55<00:00,  1.36it/
                 all       5000      36335      0.743      0.627      0.685      0.503
Speed: 0.1ms pre-process, 7.5ms inference, 2.1ms NMS per image at shape (32, 3, 640, 640)  # <--- baseline speed

Evaluating pycocotools mAP... saving runs/val/exp3/yolov5x_predictions.json...

...
Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.505 # <--- baseline mAP
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.689
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.545
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.339
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.557
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.650
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.382
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.628
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.677  # <--- baseline mAR
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.523
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.730
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.826

4

融合测试

通过在任何现有的 val.py或detect.py命令中的 –weights 参数后添加额外的模型，可以在测试和推理时将多个预训练模型融合在一起。

将 yolov5x, yolov5l6 两个模型的融合测试的指令如下：

python val.py --weights ./yolov5x ./yolov5l6  --data data/coco.yaml --img 640 --half

val: data=data/coco.yaml, weights=['./yolov5x', './yolov5l6'], batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, task=val, device=, workers=8, single_cls=False, augment=False, verbose=False, save_txt=False, save_hybrid=False, save_conf=False, save_json=True, project=runs/val, name=exp, exist_ok=False, half=True, dnn=False
YOLOv5  v1.0-29-g8ed33f3 Python-3.8.13 oneflow-0.8.1.dev20221018+cu112 
Fusing layers... 
Model summary: 322 layers, 86705005 parameters, 571965 gradients
Fusing layers... 
Model summary: 346 layers, 76726332 parameters, 653820 gradients
Ensemble created with ['./yolov5x', './yolov5l6']

val: Scanning 'https://b2.7b2.com/data/dataset/fengwen/coco/val2017.cache' images and labels... 4952 found, 48 missing, 0 empty, 0 corrupt: 100%|████████
               Class     Images     Labels          P          R     mAP@.5 mAP@.5:.95: 100%|██████████| 157/157 [03:14<00:00,  1.24s/i
                 all       5000      36335       0.73      0.644      0.693      0.513
Speed: 0.1ms pre-process, 23.7ms inference, 2.3ms NMS per image at shape (32, 3, 640, 640) # <--- ensemble speed

Evaluating pycocotools mAP... saving runs/val/exp21/yolov5x_predictions.json...

...
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.515  # <--- ensemble mAP
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.697
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.556
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.340
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.567
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.678
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.389
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.637
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.690 # <--- ensemble mAR
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.517
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.743
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.842

声明：上述两次测试的mAP，mAR结果如下：

5

融合推理

附加额外的模型在 –weights 选项后自动启用融合推理：

python detect.py --weights ./yolov5x ./yolov5l6 --img 640 --source  data/images

Output:

detect: weights=['./yolov5x', './yolov5l6'], source=data/images, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=False, dnn=False
YOLOv5  v1.0-31-g6b1387c Python-3.8.13 oneflow-0.8.1.dev20221018+cu112 
Fusing layers... 
Model summary: 322 layers, 86705005 parameters, 571965 gradients
Fusing layers... 
Model summary: 346 layers, 76726332 parameters, 653820 gradients
Ensemble created with ['./yolov5x', './yolov5l6']

detect.py:159: DeprecationWarning: In future, it will be an error for 'np.bool_' scalars to be interpreted as an index
  s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string
image 1/2 /home/fengwen/one-yolov5/data/images/bus.jpg: 640x512 4 persons, 1 bus, 1 handbag, 1 tie, Done. (0.028s)
detect.py:159: DeprecationWarning: In future, it will be an error for 'np.bool_' scalars to be interpreted as an index
  s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string
image 2/2 /home/fengwen/one-yolov5/data/images/zidane.jpg: 384x640 3 persons, 2 ties, Done. (0.023s)
0.6ms pre-process, 25.6ms inference, 2.4ms NMS per image at shape (1, 3, 640, 640)

参考文章

• github.com/ultralytics…

欢迎 Star、试用 OneFlow 最新版本：

github.com/Oneflow-Inc…