Regarding the calculation of the number of YOLOv8 detection head parameters.

[qingyun259]

Search before asking

• [X] I have searched the YOLOv8 issues and discussions and found no similar questions.

Question

I made a change to the original YOLOv8 yaml file to include P2 and build an additional detection head, but when I printed the model information I found that the number of parameters in the detection head had dropped by almost half. But shouldn't building a new detection header increase the number of parameters in this part? May I ask what causes such a result?

Additional

No response

[github-actions[bot]]

👋 Hello @qingyun259, thank you for your interest in Ultralytics YOLOv8 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the vibrant Ultralytics Discord 🎧 community for real-time conversations and collaborations. This platform offers a perfect space to inquire, showcase your work, and connect with fellow Ultralytics users.

Install

Pip install the ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8.

pip install ultralytics

Environments

YOLOv8 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLOv8 Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

@qingyun259 hello,

Thank you for reaching out with your question. When you add a new detection head to the YOLOv8 model, the expected behavior would typically be an increase in the number of parameters. However, if you're observing a decrease, it could be due to several factors:

1. Parameter Sharing: Check if the new detection head shares layers or parameters with existing heads, which might not add additional parameters as expected.
2. Configuration Settings: Ensure that the modifications in the YAML file correctly specify the layers and connections for the new detection head. A misconfiguration could lead to unintended layer reductions elsewhere.
3. Model Pruning: If any form of model pruning or optimization is active during training or setup, it might reduce the number of parameters.

I recommend reviewing the specific changes made in the YAML configuration and verifying that all layers and connections are defined as intended. Additionally, examining the model summary in detail can help pinpoint where the parameter reduction is occurring.

If the issue persists, please provide the modified YAML file or further details, and we can look deeper into this.

[qingyun259]

Thanks for your prompt reply, by checking the model structure using tensorboard, I found that the model is loaded normally as per the yaml file and at the same time I am not pruning it. So I would like to ask, how does YOLOv8's detection head do the sharing of parameters? What is known now is that there have been four scales of feature maps input to the detection head and they are 160×160×160, 320×80×80, 640×40×40, 640×20×20, is there a 'normalization' rule when the detection head fuses these feature maps for feature map matching. Most importantly, in my yaml file above, the detection head section is introducing additional P2 branches (corresponding to feature map size of 160×160×160) and halving the number of parameters is something that is very confusing to me, can you answer this? Or please tell me where to parse the source code for YOLOv8 parameter sharing.

[glenn-jocher]

Hello @qingyun259,

Thank you for the detailed follow-up and for using TensorBoard to verify the model structure. In YOLOv8, parameter sharing across detection heads typically isn't the default behavior unless explicitly configured to do so. Each detection head processes its respective scale of feature maps independently unless there's a cross-connection or a merging layer defined that explicitly shares weights.

Regarding the fusion of feature maps, YOLOv8 does not inherently normalize or standardize feature maps before they are input to the detection heads unless specified in your model configuration. The detection heads process the input feature maps as they are, applying the respective convolutional filters.

The reduction in parameters with the addition of the P2 branch might be due to several factors, including the specific configuration of convolutional layers or filters in that branch. It's possible that the configuration leads to fewer parameters being used more efficiently.

For a deeper dive into how YOLOv8 handles these configurations, I recommend looking into the source code, particularly around the modules responsible for building the detection heads and feature map processing. This can often clarify how parameters are being utilized and shared across different parts of the model.

If you need more specific guidance, feel free to share the segment of your YAML file concerning the detection heads, and I can provide more targeted advice.

[qingyun259]

`# Ultralytics YOLO 🚀, AGPL-3.0 license

YOLOv8 object detection model with P2-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect

Parameters

nc: 80 # number of classes scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'

[depth, width, max_channels]

n: [0.33, 0.25, 1024] s: [0.33, 0.50, 1024] m: [0.67, 0.75, 768] l: [1.00, 1.00, 512] x: [1.00, 1.25, 512]

YOLOv8.0 backbone

backbone:

[from, repeats, module, args]

• [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
• [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
• [-1, 3, C2f, [128, True]]
• [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
• [-1, 6, C2f, [256, True]]
• [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
• [-1, 6, C2f, [512, True]]
• [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
• [-1, 3, C2f, [1024, True]]
• [-1, 1, SPPF, [1024, 5]] # 9

YOLOv8.0-p2 head

head:

• [-1, 1, nn.Upsample, [None, 2, 'nearest']]

• [[-1, 6], 1, Concat, [1]] # cat backbone P4

• [-1, 3, C2f, [512]] # 12

• [-1, 1, nn.Upsample, [None, 2, 'nearest']]

• [[-1, 4], 1, Concat, [1]] # cat backbone P3

• [-1, 3, C2f, [256]] # 15 (P3/8-small)

• [-1, 1, nn.Upsample, [None, 2, 'nearest']]

• [[-1, 2], 1, Concat, [1]] # cat backbone P2

• [-1, 3, C2f, [128]] # 18 (P2/4-xsmall)

• [-1, 1, Conv, [128, 3, 2]]

• [[-1, 15], 1, Concat, [1]] # cat head P3

• [-1, 3, C2f, [256]] # 21 (P3/8-small)

• [-1, 1, Conv, [256, 3, 2]]

• [[-1, 12], 1, Concat, [1]] # cat head P4

• [-1, 3, C2f, [512]] # 24 (P4/16-medium)

• [-1, 1, Conv, [512, 3, 2]]

• [[-1, 9], 1, Concat, [1]] # cat head P5

• [-1, 3, C2f, [1024]] # 27 (P5/32-large)

• [[18, 21, 24, 27], 1, Detect, [nc]] # Detect(P2, P3, P4, P5) `

Thank you for your reply. The yaml file yolov8-p2.json is what I used during the training. I hope you can help me understand why the number of parameters of the detection head decreases sharply after the P2 branch is introduced.

[glenn-jocher]

Hello @qingyun259,

Thanks for sharing your YAML configuration. The sharp decrease in the number of parameters after introducing the P2 branch might be due to the specific configuration of convolutional layers or the way layers are concatenated and processed.

In your configuration, the detection head layers, especially where concatenation (Concat) and feature processing (C2f) occur, play a crucial role. When you introduce the P2 branch and its subsequent concatenation with higher-level features, it's possible that the resulting feature maps are being processed in a way that doesn't proportionally increase the number of parameters. This could be due to the efficient use of shared features across different scales or the specific settings of the C2f blocks which might be optimizing the feature channels.

To further investigate, I recommend:

• Checking the output sizes of each layer using a forward pass debug to ensure that the feature maps are as expected.
• Reviewing the C2f module's implementation to understand how it processes and combines the features, which might reveal why adding P2 doesn't increase parameters as expected.

If you need more detailed analysis, consider posting this part of your YAML and any relevant model summaries that show the parameter counts per layer. This might help in pinpointing the exact cause of the parameter reduction.

Keep exploring and tweaking, you're on the right track! 🚀

[qingyun259]

Thank you for your guidance, based on your comments I have found the possible causes which will be very positive for my research work. I wish you all the best!!!

[glenn-jocher]

Hello @qingyun259,

I'm thrilled to hear that you've found the guidance helpful and that it has positively impacted your research! If you have any more questions or need further assistance as you continue your work, feel free to reach out. Wishing you the very best in your research endeavors! 🚀

Warm

[qingyun259]

Hello, I seem to have found the reason for the parameter degradation caused by adding the P2 layer in my experiments some time ago, so I apologize for only correcting this issue now. It seems that the reason for the lowering of the parameters is not due to parameter sharing, but rather to the fact that YOLOv8 has a different "channel compression ratio" when dealing with features of different scales. This code is mainly responsible：

1. ultralytics/nn/modules/head.py→class Detect(nn.Module):...

c2, c3 = max((16, ch[0] // 4, self.reg_max * 4)), max(ch[0], min(self.nc, 100)) # channels After looking at the range of variation of the relevant channel parameters I found that the channel numbers c2 and c3's are highly dependent on ch[0], which in YOLOv8 corresponds to the number of channels in layer P3, and after the introduction of layer P2 ch[0] corresponds to the number of channels in layer P2. The shallow feature P2 has only half the number of channels of P3, so this greatly limit the number of channels corresponding to the feature map in the detection head, hence the modeled detection head position halves the number of parameters after the introduction of the P2 layer. By now, I think this is the main reason for my previous question. Finally, thank you very much for the technical support I received from you and the ultralytics team!

[glenn-jocher]

Hello @qingyun259,

Thank you for your detailed follow-up and for sharing your findings regarding the parameter degradation issue. Your investigation into the "channel compression ratio" and its impact on the detection head parameters is insightful and valuable.

Indeed, the way YOLOv8 handles feature maps at different scales can significantly influence the number of parameters, especially when introducing additional layers like P2. The code snippet you referenced from ultralytics/nn/modules/head.py highlights how the channel numbers are calculated based on the input channels, which explains the reduction in parameters when P2 is introduced.

Your explanation about the dependency of c2 and c3 on ch[0] and how it affects the detection head parameters is spot on. The shallower P2 layer having fewer channels than P3 indeed leads to a reduced number of parameters in the detection head.

If you have any further questions or need additional support, feel free to ask. We're here to help! 😊

Best of luck with your research, and thank you for your kind words about the Ultralytics team!

Warm