Passing a single layer (e.g., nn.Conv2d) resulted in 0 params

[Yui-Arthur]

Search before reporting

• I have searched the appropriate Ultralytics issues and found no similar bug report.

Repository or project

https://github.com/ultralytics/thop

Component / area

No response

Bug description

• Got the zero parameters when profile a single layer (e.g. nn.Conv2d)

Environment

• OS: Windows11 24H2
• Python: 3.12.12
• Install: git

Package Version

• filelock: 3.20.0
• fsspec: 2025.12.0
• Jinja2: 3.1.6
• MarkupSafe: 2.1.5
• mpmath: 1.3.0
• networkx: 3.6.1
• numpy: 2.3.5
• packaging: 25.0
• pillow: 12.0.0
• pip: 26.0.1
• setuptools: 80.10.2
• sympy: 1.14.0
• torch: 2.10.0+cu128
• torchvision: 0.25.0+cu128
• typing_extensions: 4.15.0
• wheel: 0.46.3

Minimal reproducible example

from thop import profile
import torch
import torch.nn as nn

class test_model(nn.Module):
    def __init__(self):
        super().__init__()
        self.layer = nn.Conv2d(3, 10, (3,3))

    def forward(self, input):
        return self.layer(input)


layer = nn.Conv2d(3, 10, (3,3))
model = test_model()

macs, weight = profile(layer, torch.zeros(1,3,224,224))
print("Single Layer Profile Weights: ", weight)
# Single Layer Profile Weights:  0
macs, weight = profile(model, torch.zeros(1,3,224,224))
print("Model Profile Weights: ", weight)
# Model Profile Weights:  280.0

Additional context

• In example, the weights of nn.Conv2d(3, 10, (3,3)) should be 280 but got 0

Are you willing to submit a PR?

• Yes, I'd like to help by submitting a PR!

[UltralyticsAssistant]

👋 Hello @Yui-Arthur, thank you for submitting a ultralytics/thop 🚀 Issue. This is an automated response to help us triage quickly, and an Ultralytics engineer will also assist soon 🛠️🙂 To help us address your concern efficiently, please ensure you've provided the following information:

1. For bug reports:

   • A clear and concise description of the bug (you mentioned profiling a single layer like nn.Conv2d returns 0 params ✅)
   • A minimum reproducible example MRE that demonstrates the issue (thanks for including this ✅)
   • Your environment details (OS, Python version, package versions) (provided ✅)
   • Expected behavior vs. actual behavior (please add what parameter count you expect vs. what you’re seeing)
   • Any error messages or logs related to the issue (if any, please paste them)

2. For feature requests:

   • A clear and concise description of the proposed feature
   • The problem this feature would solve
   • Any alternative solutions you've considered

3. For questions:

   • Provide as much context as possible about your question
   • Include any research you've already done on the topic
   • Specify which parts of the documentation, if any, you've already consulted

Please make sure you've searched existing issues to avoid duplicates. If you need to add any additional information, please comment on this issue.

[pderrenger]

Yep, I can reproduce this: in the current profile() implementation the recursive dfs_count() initializes total_params to 0 and only accumulates child modules, so when the root module is itself a leaf (e.g., nn.Conv2d), you end up with 0 params even though the layer has weights. (raw.githubusercontent.com)

As a workaround, pass inputs as a tuple and either wrap the layer or use profile_origin() (which counts params correctly for leaf roots): (github.com)

x = torch.zeros(1, 3, 224, 224)
macs, params = profile(nn.Sequential(layer), inputs=(x,))      # workaround
# or
macs, params = profile_origin(layer, inputs=(x,))              # workaround

If you’d like to PR a fix, updating dfs_count() to start with module.total_params.item() (instead of 0) should resolve single-layer profiling while keeping behavior for full models unchanged. (raw.githubusercontent.com)