v2.1.4 - Count nn.Bilinear instead of charging it nothing (#158)

🌟 Summary

THOP 2.1.4 now accurately counts nn.Bilinear operations instead of treating them as free, improving model complexity and compute reporting. 🎯

📊 Key Changes

• Added nn.Bilinear support to THOP’s operation-counting registry.
• Introduced a dedicated count_bilinear hook that accounts for both contractions performed by PyTorch.
• Counts operations across standard, batched, and higher-dimensional inputs using the output tensor size.
• Updated the package version from 2.1.3 to 2.1.4.
• Change contributed by @raimbekovm in PR #158.

🎯 Purpose & Impact

• More accurate profiling: Models using bilinear layers no longer report artificially low operation counts.
• Better comparisons: FLOPs/MACs estimates now better reflect the actual computational cost of architectures such as attention, fusion, and multimodal networks.
• Consistent behavior: The implementation follows THOP’s existing linear-layer counting convention without adding shape-specific branches.
• Improved resource planning: More reliable estimates can help users compare models and anticipate inference or training costs. 🚀

What's Changed

• Count nn.Bilinear instead of charging it nothing by @raimbekovm in #158

Full Changelog: v2.1.3...v2.1.4