sovrasov/flops-counter.pytorch

Health v2 · maintenance only

Adoption not part of the score

2982 stars · 306 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

A Python library (ptflops) that computes the theoretical number of multiply-add operations (FLOPs) and parameter counts for neural network models in PyTorch. It supports two backends (aten and pytorch) and can print per-layer computational cost breakdowns.

Use cases

• count FLOPs of a PyTorch model
• estimate computational complexity of a CNN or transformer
• count the number of parameters in a neural network
• get per-layer computational cost breakdown of a model
• compare efficiency of model architectures
• measure inference cost of a vision transformer
• profile model complexity for a research paper

When to choose

• you need theoretical FLOPs or MACs estimates for PyTorch models including transformers
• you want per-layer cost analytics for CNNs
• you need a lightweight pip-installable complexity counter with MIT license

When to avoid

• you need runtime latency or actual measured speed rather than theoretical operation counts
• you work outside PyTorch (e.g., TensorFlow, JAX)
• you need GPU memory profiling or training-time cost estimation

Facets

library · maturity active

benchmarking machine-learning deep-learning developer-tools deep-learning machine-learning performance developer-tools python cross-platform flops-counter pytorch model-complexity ptflops neural-networks transformers cnn parameter-counting

Member repositories

For agents

markdown · JSON · MCP: product_card(name="sovrasov/flops-counter.pytorch")