Replace thop with torch.profiler for FLOPs

[glenn-jocher]

@Burhan-Q as described replacement for thop. If this works well we can eliminate the dependency entirely.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Redefine how model FLOPs are calculated, enhancing compatibility and removing dependencies.

📊 Key Changes

• Deprecated the get_flops function that relied on the thop package.
• Introduced get_flops as a new method, leveraging the built-in torch.profiler for FLOPs calculation, aimed at removing external dependencies and enhancing the function's integration with newer PyTorch versions.
• Adjusted the methodology to calculate FLOPs based on either stride size or actual image size, ensuring versatility across different model architectures.

🎯 Purpose & Impact

• Purpose: Streamline and make the process of calculating a model's Floating Point Operations Per Second (FLOPs) more reliable and integrated with PyTorch ecosystems. This change aims to ensure better compatibility with future updates and reduce dependency issues.
• Impact: Developers and users working with Ultralytics models will experience a smoother, more integrated method for FLOPs calculation. This can lead to more accurate performance assessments and optimizations without relying on third-party packages.
• For Developers: Easier integration and maintenance with PyTorch's evolving features while reducing external dependency risks.
• For Users: Transparent and consistent performance metrics aiding in model selection and optimization efforts.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 66.60%. Comparing base (8fb1406) to head (c595873).

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #12634      +/-   ##
==========================================
- Coverage   70.26%   66.60%   -3.67%     
==========================================
  Files         124      124              
  Lines       15681    15660      -21     
==========================================
- Hits        11019    10431     -588     
- Misses       4662     5229     +567     

Flag	Coverage Δ
Benchmarks	35.40% <100.00%> (ø)
GPU	?
Tests	63.36% <100.00%> (-3.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[glenn-jocher]

Ok after testing this seems to be mostly working (few edge cases left, but it's MUCH SLOWER than thop. For the fastest FLOPs computations I see 0.02 thop and 0.05 torch profiler, for the very slowest (RTDETR at 640) I see 1s thop and 11s - 13s torch profiler.

This speed reduction is terrible, especially for RTDETR and possibly SAM, so I think while this was a great experiment it may have to stay an experiment for now.