v2.1.6 - Release 2.1.6 (#174)

🌟 Summary

🚀 THOP 2.1.6 is a patch release that improves model-operation profiling accuracy and robustness while preserving the existing API and behavior for typical users.

📊 Key Changes

• 📦 Released THOP 2.1.6

  • Updated the package version from 2.1.5 to 2.1.6.
  • No intentional breaking API changes.

• 🧮 More accurate recurrent-layer profiling

  • Packed RNN, GRU, and LSTM inputs are now counted using their actual processed sequence steps instead of the larger padded shape.
  • LSTM projection layers now account for proj_size, including the additional projection work and reduced recurrent width.
  • This produces more realistic operation estimates for variable-length sequences and projected LSTMs.

• 🧩 Improved custom operation support

  • profile_origin() now applies custom profiling rules to composite modules, matching the behavior of the standard profile() function.
  • Custom operations that may have nonlinear computational costs now use direct profiling instead of inaccurate linear extrapolation.

• 🧱 Better handling of container and zero-operation modules

  • Added support for containers such as ModuleList, ModuleDict, ParameterList, and ParameterDict.
  • Max-unpool layers are recognized as performing no arithmetic operations.
  • Embedding lookups are correctly treated as data gathering rather than arithmetic computation.
  • Embedding layers using max_norm emit a warning because their data-dependent renormalization is not counted.
  • EmbeddingBag remains unregistered because its reduction cost depends on the input contents.

• 🛡️ Safer total_ops handling

  • THOP now rejects modules that already use total_ops for parameters, child modules, or class-level properties.
  • This prevents profiling from silently corrupting module structure, parameter counts, or reported operation totals.

• ✅ Validation

  • The complete test suite passed: 20 tests successful.

🎯 Purpose & Impact

• 📈 More trustworthy profiling results for packed sequences, projected LSTMs, custom modules, and complex model containers.
• ⚙️ Fewer false warnings when profiling models that use common PyTorch containers or embedding layers.
• 🧠 Better support for advanced architectures with custom operation rules and nonlinear computation patterns.
• 🔒 Improved reliability and error reporting when a model conflicts with THOP’s internal total_ops bookkeeping.
• 👍 For most users, this is a drop-in upgrade with no code changes required.

What's Changed

• Count a packed sequence by its real steps and read the LSTM projection by @raimbekovm in #160
• Let profile_origin count a composite module's own rule by @raimbekovm in #168
• Register the container and max-unpool types that compute nothing by @raimbekovm in #161
• Stop extrapolating a quadratic custom_ops rule as if it were affine by @raimbekovm in #167
• Register the embedding lookup and record why the bag is not counted by @raimbekovm in #165
• Reject a total_ops name a module already owns instead of misreading it by @raimbekovm in #171
• Release 2.1.6 by @glenn-jocher in #174

Full Changelog: v2.1.5...v2.1.6