v2.1.1 - Add stride-aware image profiling (#149)

🌟 Summary

🚀 THOP v2.1.1 adds stride-aware image profiling, enabling faster and more reliable MAC estimates for image models while modernizing benchmarks and documentation.

📊 Key Changes

• Stride-aware profile() support by @glenn-jocher:
  • Adds an optional stride argument for image models.
  • Profiles smaller, stride-aligned inputs instead of always processing the full target image.
  • Uses one profile for spatial-only models and two profiles when fixed-cost layers—such as linear or adaptive pooling layers—require more accurate estimation.
  • Falls back to the original target input if proxy inputs are unsupported or unsuitable.
  • Existing calls such as profile(model, inputs=...) remain unchanged.
• Simplified Ultralytics integration:
  • Removes the need for Ultralytics-side image-size scaling and layer-inspection logic.
  • Moves profiling behavior into THOP, the component that owns the operation-counting process.
  • Preserves compatibility with YOLOv3, YOLOv5, and third-party users of the existing API.
• Version update:
  • Bumps THOP to 2.1.1 and ultralytics-thop from 2.1.0 to 2.1.1.
• Modernized benchmarks:
  • Replaces legacy torchvision benchmark tables with fused YOLOv8, YOLO11, and YOLO26 results.
  • Updates the benchmark script to reproduce 15 model variants at 640 × 640 without downloading weights.
• Improved documentation:
  • Simplifies the English README with clearer installation and usage examples.
  • Adds a Simplified Chinese README.
  • Documents the new stride profiling workflow and custom operation rules.
• Repository workflow updates:
  • Refreshes AGENTS.md with stronger guidance around minimal changes, solving problems in the owning code path, deleting duplication, and avoiding regressions.
• Validation:
  • All 20 tests passed, with Ruff checks and formatting completed successfully ✅

🎯 Purpose & Impact

• ⚡ Faster profiling: Large image models can be analyzed using much smaller stride-aligned inputs, reducing profiling overhead.
• 🎯 More accurate estimates: Two-point profiling improves results for models that combine spatially scaling operations with fixed-cost layers.
• 🧩 Less duplicated code: Consumers no longer need to maintain their own FLOPs scaling and fallback policies.
• 🔒 Better compatibility: The new behavior is opt-in, so existing applications should continue to produce the same results when stride is omitted.
• 📈 More relevant comparisons: Updated YOLOv8, YOLO11, and YOLO26 benchmarks make it easier to compare modern Ultralytics model complexity before training or deployment.
• 🌍 Broader accessibility: Clearer documentation and Chinese-language support make THOP easier to adopt for a wider developer community.

What's Changed

• Refresh README and YOLO benchmarks by @glenn-jocher in #142
• Add stride-aware image profiling by @glenn-jocher in #149

Full Changelog: v2.1.0...v2.1.1