YOLOV8 classification problem

0zxq0 · November 27, 2025, 5:43am

I am a beginner, and I am currently using YOLOv8 for classification tasks. I hope that the results of each class in the model I trained do not affect each other during detection, meaning that the probabilities are calculated separately, and their sum cannot equal one. Through searching for information, I have learned that I need to change the activation function from Softmax to sigmoid. However, I am unsure where to start. I hope someone can help and guide me on how to achieve this task

Toxite · November 27, 2025, 5:44am

You need to change this line to x.sigmoid()

github.com/ultralytics/ultralytics

ultralytics/nn/modules/head.py

0f9fa1f58


      
                  self.drop = nn.Dropout(p=0.0, inplace=True)
                  self.linear = nn.Linear(c_, c2)  # to x(b,c2)
          
              def forward(self, x: list[torch.Tensor] | torch.Tensor) -> torch.Tensor | tuple:
                  """Perform forward pass of the YOLO model on input image data."""
                  if isinstance(x, list):
                      x = torch.cat(x, 1)
                  x = self.linear(self.drop(self.pool(self.conv(x)).flatten(1)))
                  if self.training:
                      return x
                  y = x.softmax(1)  # get final output
                  return y if self.export else (y, x)
          
          
          class WorldDetect(Detect):
              """Head for integrating YOLO detection models with semantic understanding from text embeddings.
          
              This class extends the standard Detect head to incorporate text embeddings for enhanced semantic understanding in
              object detection tasks.
          
              Attributes:

0zxq0 · November 27, 2025, 6:18am

Thank you very much for your response

pderrenger · November 28, 2025, 12:38am

Glad that helped!

One small clarification that may be useful as you keep experimenting: changing that line in the classification head from Softmax to Sigmoid only affects the inference output, not how the model is trained. During training, YOLO’s classification head returns raw logits and the loss is still standard cross‑entropy (single-label, classes compete). The Softmax or Sigmoid is only applied after training for easier probability interpretation.

If you truly need independent multi‑label probabilities (an image can belong to several classes at once), you would also need to:

change the loss to something like BCEWithLogitsLoss
prepare labels as multi‑hot vectors (multiple 1s per sample)

That is more involved and not fully plug‑and‑play in the current classification pipeline.

For a quick conceptual refresher on what you just changed, the articles on the Sigmoid function and the Softmax activation explain the difference in more detail.

0zxq0 · November 28, 2025, 12:53am

Thank you very much. I have replaced the loss function.

Topic		Replies	Views
Yolov8 architecture modifying for multi-label classification YOLO yolo , question , support , discussion , code	3	348	November 21, 2025
Loss function in classification yolov8 YOLO yolov8 , troubleshooting	3	1419	November 27, 2024
Add New Classes to (YOLOv8n or YOLO11n) Pretrained Model Without Losing COCO Classes Discussion yolo , question , support	2	1082	April 20, 2025
YOLOv11 uses multiple heads(classification+segmentation) during training Discussion yolo , question , code	1	132	December 11, 2025
New Release: Ultralytics v8.3.47 Discussion releases , announcements , ultralytics-official	0	63	December 7, 2024

YOLOV8 classification problem

Related topics