A profiler for matrix math, not milliseconds

What it does THOP profiles PyTorch models by tallying multiply-accumulate operations (MACs), floating-point operations (FLOPs), and trainable parameters for a given input shape. It also publishes a reference table of classic torchvision CNNs—from AlexNet to ShuffleNet—so you can stack your own model against familiar baselines.

The interesting bit The library exposes a custom_ops hook that lets you define counting rules for third-party or bespoke modules, preventing oddball layers from silently dropping out of the total. It also includes a clever_format utility to turn raw integer counts into readable strings, because staring at ten-digit numbers is nobody’s idea of fun.

Key highlights

• Returns both MACs/FLOPs and parameter count from a single profile call.
• Supports custom counting functions for unsupported modules via a dictionary hook.
• Formats output cleanly with clever_format instead of dumping raw integers.
• Bundles a benchmark script and results table for standard CNNs adapted from torchvision.

Verdict Useful if you need a quick, architecture-level sanity check on model complexity. Skip it if your work is already covered by a lower-level profiler that traces kernel execution times.