YOLOv8 Final Detection Head Still Outputs (1, 7, 8400) Instead of (1, 8, 8400) for 3 Classes

[Saleh-1987]

I trained a YOLOv8 detection model with 3 classes, but the raw forward pass still shows a final detect output of (1, 7, 8400) instead of (1, 8, 8400).

What I’ve Done:
Checked my data.yaml:

yaml
train: path/to/train/images
val: path/to/val/images
nc: 3
names: ['glioma', 'meningioma', 'pituitary']

Confirmed nc: 3 is correct.
Trained from scratch with the command:

bash
yolo detect train \
    data=path/to/data.yaml \
    model=yolov8x \
    epochs=1000 \
    imgsz=640 \
    device=1 \
    patience=100

The training runs without error and completes successfully.
Installed the latest Ultralytics version (v8.3.72) to ensure no version issues:

bash
pip uninstall ultralytics
pip install ultralytics

Loaded the new best.pt directly:

python
from ultralytics import YOLO
import torch

model = YOLO(r"best.pt").model
model.eval()

dummy_input = torch.randn(1, 3, 640, 640)
with torch.no_grad():
    outputs = model(dummy_input)

for out in outputs:
    # Some outputs are lists; checking each element carefully
    if isinstance(out, torch.Tensor):
        print(out.shape)
    else:
        print("List output:", [o.shape for o in out if hasattr(o, 'shape')])

The console shows (1, 7, 8400) for the detection output.
Verified model metadata says nc=3 and model.names has 3 classes. However, the raw detect layer output is still 7 channels.

Observations:
If a YOLO detect layer is genuinely for 3 classes, it should output (5 + 3)=8 channels per anchor, not 7.
The mismatch (1, 7, 8400) typically indicates it’s still set for 2 classes despite nc=3.

Question / Request for Help:
Why is the raw detect head still (1, 7, 8400) even though I trained from scratch for 3 classes?
How can I ensure the detect layer is fully re-initialized to (5 + 3)=8 for 3-class detection?
I’ve tried deleting old .pt files, re-checking my data.yaml, reinstalling ultralytics, and confirming model.model.nc == 3. But the final detect layer continues to yield 7 channels instead of 8.
Any ideas on what might cause this persistent mismatch?

Thanks in advance for any help or insights!

[UltralyticsAssistant]

👋 Hello @Saleh-1987, thank you for bringing this to our attention and for your detailed explanation 🚀! It seems like you’re encountering an issue with the detection head output not aligning with the number of classes defined during training.

We recommend double-checking a few key areas to help diagnose the problem, and here’s how you can proceed:

• 🐛 Bug Report? If you suspect this is a bug, please provide a Minimum Reproducible Example (MRE) that we can use to investigate further. This will help us replicate the behavior on our side and determine the root cause.

• 🔧 Custom Training? For custom training concerns, ensure the following:

  1. Your data.yaml file is correctly configured, as you’ve shown above.
  2. The training logs and loss functions are behaving as expected (e.g., no unusual spikes or rapid collapses).
  3. Follow our Tips for Best Training Results to ensure optimal configuration.

• 📦 Upgrade: Confirm you are using the latest ultralytics version (v8.3.72 as you’ve stated). You can double-check this with:

  pip show ultralytics

  If needed, upgrade using:

  pip install -U ultralytics

• ✅ Validation Checklist:

  1. Verify the loaded best.pt model reflects nc=3. You can check this with:

     from ultralytics import YOLO
     model = YOLO('best.pt').model
     print(model.nc, model.names)

  2. Confirm the detect layer structure is correct:

     print(model.model[-1])  # Inspect final detection layer

  3. Ensure to remove any cached files that might conflict, particularly if training was restarted multiple times.

If the issue persists, you might also want to reset and reinitialize your train scripts or checkpoints to ensure no residual configurations are affecting the outputs.

---

Community Support 🌟

You can also engage with the Ultralytics community for insights and ideas:

• Discord 🎧 for real-time discussion.
• Discourse for in-depth assistance.
• Subreddit to exchange ideas with peers.

Verified Environments

Ensure you’re working within a verified environment for YOLO. Options include:

• Notebooks: Use cloud environments with free GPUs to validate the model setup:
  
  
  
• Docker: Preconfigured Docker Image .

---

This is an automated response 🤖, but rest assured that an Ultralytics engineer will review your issue and provide additional assistance soon! Thank you for your patience and for contributing to the Ultralytics community ❤️.

[glenn-jocher]

It appears that the updated class count wasn’t properly applied during model construction, so please ensure the training configuration (including nc) is correctly used when building and saving the model.

[Y-T-G]

4 coordinates + 3 class scores = 7

It's correct.

[Saleh-1987]

Yes, that's correct as YOLOv8 does not include an explicit objectness score. Instead, class probabilities already incorporate object confidence.

[glenn-jocher]

The output shape (1,7,8400) is correct for YOLOv8 with 3 classes. YOLOv8 uses 4 coordinates (x,y,w,h) + 3 class scores (which implicitly include objectness confidence), totaling 7 channels. This is an intentional architectural design - there's no separate objectness score. Your model is working as expected. For more details, see our detection head documentation at https://docs.ultralytics.com/reference/nn/modules/head/#ultralytics.nn.modules.head.Detect

[Saleh-1987]

YOLOv8 does not include an explicit objectness score. Instead, class probabilities already incorporate object confidence. Hence, the (1, 7, 8400) is correct.

[glenn-jocher]

@Saleh-1987 thanks for reaching out! 😊 The (1,7,8400) output in YOLOv8 is expected because it uses class probabilities that inherently include object confidence, eliminating the need for a separate objectness score. For 3 classes, the channels are calculated as 4 (bbox) + 3 (class scores) = 7. This is normal behavior and indicates your model is functioning correctly. For more details on the detection head architecture, check out our Detect module docs: https://docs.ultralytics.com/reference/nn/modules/head/. Let us know if you have further questions!