PyTorch model autopsy: count flops without the guesswork

What it does

torchstat hooks into a PyTorch nn.Module and prints a per-layer table of theoretical compute costs: FLOPs, multiply-adds, parameter counts, and memory read/write estimates. You can call it from the CLI against a Python file or import it as a module and pass your model directly.

The interesting bit

The tool tries to estimate memory traffic (MemRead/MemWrite) alongside arithmetic, which most model summaries skip. That makes it slightly more useful for reasoning about bandwidth-bound vs. compute-bound layers, though the README warns that some layers like Dropout2d simply aren’t supported yet.

Key highlights

• CLI and Python API both available; stat(model, (3, 224, 224)) is the one-liner
• Reports per-layer and total: params, FLOPs, MAdd, memory in MB, and byte-level memory traffic
• Layer support is limited to nn.Module subclasses; torch.nn.functional ops are explicitly not supported
• Built on prior work from flops-counter.pytorch and pytorch_model_summary
• Requires PyTorch 0.4.0+ and Python 3.6+ (versions that feel increasingly historical)

Caveats

• README marks the repo as “currently under development” with API changes possible
• Several TODOs remain un-checked: detailed layer-wise summary, export to table, and arbitrary input shapes
• Some layers emit blunt “not supported” warnings and report zero cost rather than failing gracefully

Verdict

Worth a look if you need quick, static estimates of model complexity on older PyTorch codebases. Skip it if you need production-grade profiling, dynamic shapes, or functional API support; PyTorch’s built-in profiler and tools like deepspeed.profiling have largely superseded this niche.