thop 0.0.23-1908021522

A tool to count the FLOPs of PyTorch model.

THOP: PyTorch-OpCounter

How to install

• Through PyPi

  pip install thop

• Using GitHub (always latest)

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

How to use

• Basic usage

  from torchvision.models import resnet50
  from 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 definition
  def count_your_model(model, x, y):
      # your rule here
  
  input = 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 on Recent Models

Model Params(M) FLOPs(G) alexnet 58.27 0.77 densenet121 7.61 2.71 densenet161 27.35 7.34 densenet169 13.49 3.22 densenet201 19.09 4.11 resnet101 42.49 7.34 resnet152 57.40 10.82 resnet18 11.15 1.70 resnet34 20.79 3.43 resnet50 24.37 3.86

Model Params(M) FLOPs(G) squeezenet1_0 1.19 1.01 squeezenet1_1 1.18 0.48 vgg11 126.71 7.98 vgg11_bn 126.71 8.01 vgg13 126.88 11.82 vgg13_bn 126.89 11.86 vgg16 131.95 16.12 vgg16_bn 131.96 16.17 vgg19 137.01 20.43 vgg19_bn 137.02 20.49