前一篇文章详细了讲解了如何构造自己的数据集，以及如何修改模型配置文件和数据集配置文件，本篇主要是如何训练自己的数据集，并且如何验证。
VOC2012数据集下载地址：
http://host.robots.ox.ac.uk/pascal/VOC/voc2012/
coco全量数据集下载地址：
http://images.cocodtaset.org/annotations/annotations_trainval2017.zip
本篇以以下图片为预测对象。
一、对coco128数据集进行训练，coco128.yaml中已包括下载脚本，选择yolov8n轻量模型，开始训练
yolo detect train data=coco128.yaml model=model\yolov8n.pt epochs=100 imgsz=640
训练的相关截图，第一部分是展开后的命令行执行参数和网络结构
第二部分是每轮训练过程
第三部分是对各类标签的验证情况
二、对VOC2012数据集进行训练，使用我们定义的两个yaml配置文件，选择yolov8n轻量模型，开始训练
yolo detect train data=E:\JetBrains\PycharmProject\Yolov8Project\venv\Lib\site-packages\ultralytics\cfg\datasets\VOC2012.yaml model=E:\JetBrains\PycharmProject\Yolov8Project\venv\Lib\site-packages\ultralytics\cfg\models\v8\VOC2012.yaml pretrained=model\yolov8n.pt epochs=10 imgsz=640
以下为运行日志，和上述一样
(venv) PS E:\JetBrains\PycharmProject\Yolov8Project> yolo detect train data=E:\JetBrains\PycharmProject\Yolov8Project\venv\Lib\site-packages\ultralytics\cfg\datasets\VOC2012.yaml model=E:\JetBrains\PycharmProject\Yolov8Project\venv\
Lib\site-packages\ultralytics\cfg\models\v8\VOC2012.yaml pretrained=model\yolov8n.pt epochs=10 imgsz=640
WARNING no model scale passed. Assuming scale='n'.
from n params module arguments
0-11464 ultralytics.nn.modules.conv.Conv[3, 16, 3, 2]
1-114672 ultralytics.nn.modules.conv.Conv[16, 32, 3, 2]
2-117360 ultralytics.nn.modules.block.C2f [32, 32, 1, True]
3-1118560 ultralytics.nn.modules.conv.Conv[32, 64, 3, 2]
4-1249664 ultralytics.nn.modules.block.C2f [64, 64, 2, True]
5-1173984 ultralytics.nn.modules.conv.Conv[64, 128, 3, 2]
6-12197632 ultralytics.nn.modules.block.C2f [128, 128, 2, True]
7-11295424 ultralytics.nn.modules.conv.Conv[128, 256, 3, 2]
8-11460288 ultralytics.nn.modules.block.C2f [256, 256, 1, True]
9-11164608 ultralytics.nn.modules.block.SPPF [256, 256, 5]
10-110 torch.nn.modules.upsampling.Upsample[None, 2, 'nearest']
11[-1, 6] 10 ultralytics.nn.modules.conv.Concat[1]
12-11148224 ultralytics.nn.modules.block.C2f [384, 128, 1]
13-110 torch.nn.modules.upsampling.Upsample[None, 2, 'nearest']
14[-1, 4] 10 ultralytics.nn.modules.conv.Concat[1]
15-1137248 ultralytics.nn.modules.block.C2f [192, 64, 1]
16-1136992 ultralytics.nn.modules.conv.Conv[64, 64, 3, 2]
17[-1, 12] 10 ultralytics.nn.modules.conv.Concat[1]
18-11123648 ultralytics.nn.modules.block.C2f [192, 128, 1]
19-11147712 ultralytics.nn.modules.conv.Conv[128, 128, 3, 2]
20[-1, 9] 10 ultralytics.nn.modules.conv.Concat[1]
21-11493056 ultralytics.nn.modules.block.C2f [384, 256, 1]
22[15, 18, 21] 1755212 ultralytics.nn.modules.head.Detect[20, [64, 128, 256]]
VOC2012 summary: 225 layers, 3014748 parameters, 3014732 gradients
Transferred319/355 items from pretrained weights
UltralyticsYOLOv8.0.178Python-3.10.11 torch-2.0.1+cu118 CUDA:0(Quadro P2200, 5120MiB)
engine\trainer: task=detect, mode=train, model=E:\JetBrains\PycharmProject\Yolov8Project\venv\Lib\site-packages\ultralytics\cfg\models\v8\VOC2012.yaml, data=E:\JetBrains\PycharmProject\Yolov8Project\venv\Lib\site-packages\ultralytic
s\cfg\datasets\VOC2012.yaml, epochs=10, patience=50, batch=16, imgsz=640, save=True, save_period=-1, cache=False, device=None, workers=8, project=None, name=None, exist_ok=False, pretrained=model\yolov8n.pt, optimizer=auto, verbose=
True, seed=0, deterministic=True, single_cls=False, rect=False, cos_lr=False, close_mosaic=10, resume=False, amp=True, fraction=1.0, profile=False, freeze=None, overlap_mask=True, mask_ratio=4, dropout=0.0, val=True, split=val, save
_json=False, save_hybrid=False, conf=None, iou=0.7, max_det=300, half=False, dnn=False, plots=True, source=None, show=False, save_txt=False, save_conf=False, save_crop=False, show_labels=True, show_conf=True, vid_stride=1, stream_bu
ffer=False, line_width=None, visualize=False, augment=False, agnostic_nms=False, classes=None, retina_masks=False, boxes=True, format=torchscript, keras=False, optimize=False, int8=False, dynamic=False, simplify=False, opset=None, w
orkspace=4, nms=False, lr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005, warmup_epochs=3.0, warmup_momentum=0.8, warmup_bias_lr=0.1, box=7.5, cls=0.5, dfl=1.5, pose=12.0, kobj=1.0, label_smoothing=0.0, nbs=64, hsv_h=0.015, hs
v_s=0.7, hsv_v=0.4, degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, flipud=0.0, fliplr=0.5, mosaic=1.0, mixup=0.0, copy_paste=0.0, cfg=None, tracker=botsort.yaml, save_dir=runs\detect\train8
WARNING no model scale passed. Assuming scale='n'.
from n params module arguments
0-11464 ultralytics.nn.modules.conv.Conv[3, 16, 3, 2]
1-114672 ultralytics.nn.modules.conv.Conv[16, 32, 3, 2]
2-117360 ultralytics.nn.modules.block.C2f [32, 32, 1, True]
train: Scanning E:\JetBrains\PyCharm Project\ObjectDetectionProject\datasets\VOC2012\labels\train.cache... 17125 images, 195 backgrounds, 0 corrupt: 100%|██████████| 17125/17125[00:00<?, ?it/s]
val: Scanning E:\JetBrains\PyCharm Project\ObjectDetectionProject\datasets\VOC2012\labels\train.cache... 17125 images, 195 backgrounds, 0 corrupt: 100%|██████████| 17125/17125[00:00<?, ?it/s]
Plotting labels to runs\detect\train8\labels.jpg...
optimizer: 'optimizer=auto' found, ignoring 'lr0=0.01' and 'momentum=0.937' and determining best 'optimizer', 'lr0' and 'momentum' automatically...
optimizer: AdamW(lr=0.000417, momentum=0.9) with parameter groups 57 weight(decay=0.0), 64 weight(decay=0.0005), 63 bias(decay=0.0)
Image sizes 640 train, 640 val
Using8 dataloader workers
Logging results to runs\detect\train8
Starting training for10 epochs...
Closing dataloader mosaic
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
1/102.41G0.91562.5721.24410640: 100%|██████████| 1071/1071[07:06<00:00, 2.51it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:44<00:00, 3.26it/s]
all 17125349130.6210.5720.6050.436
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
2/102.53G1.0061.8691.31110640: 100%|██████████| 1071/1071[07:06<00:00, 2.51it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:40<00:00, 3.35it/s]
all 17125349130.6440.540.5920.414
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
3/102.49G1.0381.6611.3449640: 100%|██████████| 1071/1071[07:02<00:00, 2.54it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:44<00:00, 3.25it/s]
all 17125349130.6160.5620.5940.419
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
4/102.47G1.0211.4931.33112640: 100%|██████████| 1071/1071[07:00<00:00, 2.55it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:42<00:00, 3.29it/s]
all 17125349130.6510.5880.6380.457
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
5/102.48G1.0051.4031.3184640: 100%|██████████| 1071/1071[07:00<00:00, 2.54it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:41<00:00, 3.31it/s]
all 17125349130.6730.5920.650.467
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
6/102.46G0.96821.2991.299640: 100%|██████████| 1071/1071[06:55<00:00, 2.58it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:29<00:00, 3.58it/s]
all 17125349130.7090.6230.6930.511
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
7/102.48G0.9321.2091.2618640: 100%|██████████| 1071/1071[06:57<00:00, 2.56it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:39<00:00, 3.37it/s]
all 17125349130.7210.6610.7220.542
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
8/102.49G0.89611.1271.2329640: 100%|██████████| 1071/1071[07:00<00:00, 2.55it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:40<00:00, 3.35it/s]
all 17125349130.7350.670.7460.567
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
9/102.47G0.85651.0581.2028640: 100%|██████████| 1071/1071[06:58<00:00, 2.56it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:29<00:00, 3.59it/s]
all 17125349130.7660.6960.7730.597
Epoch GPU_mem box_loss cls_loss dfl_loss InstancesSize
10/102.45G0.82780.98891.17911640: 100%|██████████| 1071/1071[06:55<00:00, 2.58it/s]
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:28<00:00, 3.61it/s]
all 17125349130.7770.7180.7950.621
10 epochs completed in 1.620 hours.
Optimizer stripped from runs\detect\train8\weights\last.pt, 6.2MB
Optimizer stripped from runs\detect\train8\weights\best.pt, 6.2MB
Validating runs\detect\train8\weights\best.pt...
UltralyticsYOLOv8.0.178Python-3.10.11 torch-2.0.1+cu118 CUDA:0(Quadro P2200, 5120MiB)
VOC2012 summary (fused): 168 layers, 3009548 parameters, 0 gradients
ClassImagesInstancesBox(P R mAP50 mAP50-95): 100%|██████████| 536/536[02:31<00:00, 3.54it/s]
all 17125349130.7770.7180.7950.621
aeroplane 171259110.9240.8130.9020.731
bicycle 171257530.7650.5780.7370.582
bird 1712511690.8940.7570.8620.651
boat 171259020.7560.6410.7260.506
bottle 1712513290.7230.5940.6790.489
bus 171256380.8930.8180.8940.775
car 1712521050.7860.690.7990.618
cat 1712512660.8520.880.9210.763
chair 1712524430.7060.5610.660.482
cow 171256420.7820.8040.8580.673
diningtable 171256350.5910.7180.690.517
dog 1712515710.8460.7950.8830.727
horse 171257600.6730.6340.740.61
person 17125157530.790.8390.8750.691
pottedplant 1712510550.7010.5250.6140.404
sheep 171258780.7750.8230.8580.665
sofa 171255920.7030.6440.730.592
train 171256720.8820.8440.9140.735
tvmonitor 171258390.730.6770.7650.595
Speed: 0.2ms preprocess, 3.9ms inference, 0.0ms loss, 0.7ms postprocess per image
Results saved to runs\detect\train8
Learn more at https://docs.ultralytics.com/modes/train
(venv) PS E:\JetBrains\PycharmProject\Yolov8Project>
三、将run\detect\trainx\best.pt拷贝到model目录下，并改为相关可辨识的模型名称
四、执行测试代码，验证一下几个训练模型的预测结果
from ultralytics import YOLO
from PIL importImage
filepath='test\eat.png'
# 直接加载预训练模型
model = YOLO('model\yolov8x.pt')
# Run inference on 'bus.jpg'
results = model(filepath) # results list
# Show the results
for r in results:
im_array = r.plot() # plot a BGR numpy array of predictions
im = Image.fromarray(im_array[..., ::-1]) # RGB PIL image
im.show() # show image
im.save('yolov8x.jpg') # save image
# 直接加载预训练模型
model = YOLO('model\yolov8n.pt')
# Run inference on 'bus.jpg'
results = model(filepath) # results list
# Show the results
for r in results:
im_array = r.plot() # plot a BGR numpy array of predictions
im = Image.fromarray(im_array[..., ::-1]) # RGB PIL image
im.show() # show image
im.save('yolov8n.jpg') # save image
# 直接加载预训练模型
model = YOLO('model\coco128.pt')
# Run inference on 'bus.jpg'
results = model(filepath) # results list
# Show the results
for r in results:
im_array = r.plot() # plot a BGR numpy array of predictions
im = Image.fromarray(im_array[..., ::-1]) # RGB PIL image
im.show() # show image
im.save('coco128.jpg') # save image
# 直接加载预训练模型
model = YOLO('model\VOC2012.pt')
# Run inference on 'bus.jpg'
results = model(filepath) # results list
# Show the results
for r in results:
im_array = r.plot() # plot a BGR numpy array of predictions
im = Image.fromarray(im_array[..., ::-1]) # RGB PIL image
im.show() # show image
im.save('VOC2012.jpg') # save image
基于yolov8x.pt预训练模型预测情况如下：
基于yolov8n.pt预训练模型预测情况如下：
基于coco128数据集训练的模型预测情况如下：
基于VOC2012数据集训练的模型预测情况如下：
结论：
1、基于yolov8x.pt预训练模型预测的最全最准，但也最慢。
2、基于yolov8n.pt预训练模型预测和yolov8x在概率上有些不一致，80类中的极少数类别识别不出来，毕竟网络模型参数相差太多。
3、基于coco128数据集训练的模型预测类别比yolov8n少，毕竟只有128张训练样本，估计会缺失一些标签。
4、基于VOC2012数据集训练的模型预测类别最少，毕竟只有20种类别，和coco数据集有交叉也有不同，VOC2012数据集没有水果样本，所以无法识别出水果。
基本上后边就可以愉快的训练各种目标检测了，但是数据集和标注数据才是比较耗人的。
最后欢迎关注公众号：python与大数据分析
