Val_batch_labels are not in the right place

Support YOLO
1

hi guys , i have encountered a problem that my val_batch_pred is giving true result while val_batch_labels is not.

val_batch0_labels
val_batch0_labels1920×460 140 KB

2

and also this is my val_batch_pred :

val_batch0_pred
val_batch0_pred1920×460 177 KB

3

This looks like the ground truth annotations are not correct. Have you verified that they use the correct YOLO format? Object Detection Datasets Overview - Ultralytics YOLO Docs

4

yes they seem to be true. here is my training code:

def main():
    torch.cuda.set_device(0)
    model = YOLO("yolov8n.yaml").load("yolov8n.pt")
    results = model.train(data='config.yaml', epochs=250, imgsz=640, device='gpu', batch=64, dropout=0.2, weight_decay=0.002, lr0=0.001, lrf=0.001)
5

The val_batch_labels are the labels from your ground truth annotations. I just tested with the coco8.yaml dataset and the labels are okay

image
image957×958 148 KB

which likely means there is no issue with the annotation plotting. This would mean that the val_batch_labels are likely showing you that your ground truth annotations are incorrect, and need to be fixed.

You can perform the same test I did with the following code:

from ultralytics import YOLO

model = YOLO("yolov8n.pt")
result = model.train(data="coco8.yaml", epochs=3)

Then navigate to the runs/detect/train directory shown at the end of training and inspect the val_batch_labels.png (there will only be one). If it doesn’t look like the one above, then you should consider upgrading you version of ultralytics as there may have been a bug in the version you have. My test was run using ultralyitcs==8.2.63

1 Like
6

thank you so much!

1 Like
7

are val_batch_labels the exact labels that i have put on the images for validation?

8

Yes, val_batch_labels.jpg are the labels for the data in images/val and labels/val where val_batch_pred.jpg are the model predictions

1 Like