torchprofile

Common errors

• ModuleNotFoundError: No module named 'torchprofile'

  cause The `torchprofile` library is not installed in the current Python environment.
  fix `pip install torchprofile`

• RuntimeError: Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same

  cause The input tensor and the model are on different devices (e.g., CPU vs. GPU). `torchprofile` requires them to be consistent.
  fix Ensure both the model and the input tensor are on the same device before profiling, for example: `model.to('cuda')` and `inputs.to('cuda')`.

• TypeError: object of type <class '...'> has no len()

  cause `torchprofile` encountered a custom layer or operation within your model for which it does not have a predefined handler, causing a failure during graph traversal.
  fix You likely need to implement and register a custom MACs handler for that specific module type. Refer to `torchprofile.handlers` for examples and instructions on `register_macs_handler`.

Warnings

• gotcha torchprofile strictly counts Multiply-Accumulate Operations (MACs). Be aware that different communities or tools might define FLOPs differently (e.g., 1 MAC = 2 FLOPs), leading to discrepancies if comparing counts across tools.
  Fix: Understand the definition used by torchprofile and adjust comparisons accordingly, or consult the source for exact definitions. The library's focus is on MACs.
• gotcha torchprofile might not automatically count operations for highly custom or non-standard `nn.Module` implementations. This can lead to underestimated MACs for models containing such layers.
  Fix: For custom modules, you may need to implement and register a specific MACs handler using `torchprofile.register_macs_handler`. Refer to the `torchprofile.handlers` module and documentation for examples.
• breaking As a 0.x version library, torchprofile's API is subject to change without strict semantic versioning. Updates, especially between minor versions (e.g., 0.0.x to 0.1.x), might introduce breaking changes.
  Fix: Pin your dependency version (e.g., `torchprofile==0.1.0`) in your project and review release notes or the GitHub changelog when upgrading to new versions.

Install

• pip install torchprofile Install stable version

Imports

• profile_macs

  from torchprofile import profile_macs

Quickstart

This quickstart demonstrates how to initialize a standard PyTorch model and use `torchprofile.profile_macs` to count its Multiply-Accumulate Operations (MACs) for a given input tensor. The result is printed in GigaMACs.

import torch
from torchvision.models import resnet18
from torchprofile import profile_macs

# Create a sample PyTorch model
model = resnet18()

# Define a dummy input tensor (batch_size, channels, height, width)
# Ensure the input device matches the model's device if applicable
inputs = torch.randn(1, 3, 224, 224)

# Profile the model MACs
macs = profile_macs(model, inputs)

print(f"ResNet-18 MACs: {macs / 1e9:.2f} G")