THOP: PyTorch-OpCounter

How to install

pip install thop (now continously intergrated on Github actions)

OR

pip install --upgrade git+https://github.com/Lyken17/pytorch-OpCounter.git

How to use

• Basic usage

  from torchvision.models import resnet50from thop import profile
  model = resnet50()input = torch.randn(1, 3, 224, 224)
  flops, params = profile(model, inputs=(input, ))

• Define the rule for 3rd party module.

  class YourModule(nn.Module):    # your definitiondef count_your_model(model, x, y):    # your rule hereinput = torch.randn(1, 3, 224, 224)
  flops, params = profile(model, inputs=(input, ), 
                          custom_ops={YourModule: count_your_model})

• Improve the output readability

  Call thop.clever_format to give a better format of the output.

  from thop import clever_format
  flops, params = clever_format([flops, params], "%.3f")

Results of Recent Models

The implementation are adapted from torchvision. Following results can be obtained using benchmark/evaluate_famours_models.py.