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:

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:

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:

pip uninstall ultralytics
pip install ultralytics

Loaded the new best.pt directly:

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?