lartpang / methodscmp

MethodsCmp: A Simple Toolkit for Counting the FLOPs/MACs, Parameters and FPS of Pytorch-based Methods

A simple toolkit for counting the FLOPs/MACs, Parameters and FPS of the model.

$ python main.py --help
usage: main.py [-h] [--method-names METHOD_NAMES [METHOD_NAMES ...]]
               [--mode {ops_params,fps,gpu_mem} [{ops_params,fps,gpu_mem} ...]]

A simple toolkit for counting the FLOPs/MACs, Parameters, FPS and GPU Memory of the model.

optional arguments:
  -h, --help            show this help message and exit
  --method-names METHOD_NAMES [METHOD_NAMES ...]
                        The names of the methods you want to evaluate.
  --mode {ops_params,fps,gpu_mem} [{ops_params,fps,gpu_mem} ...]

NOTE:

• fvcore cannot count the FLOPs of the built-in module nn.MultiHeadAttention of PyTorch

Usage

pip install fvcore

In methods folder, I have provided some recent methods, i.e. CoNet, DANet, HDFNet, JL-DCF, and UC-Net, as the examples.

More functional improvements and suggestions are welcome.

An example:

python main.py --method-names zoomnet ugtr c2fnet ujsc pfnet mgl_r slsr sinet

Change Log

• 2023-07-28:
  • [New & Important] Update the library for count FLOPs/MACS from pytorch-OpCounter to fvcore which can count FLOPs/MACs of the complex module, like Transformer.
  • Remove the useless test folder.
  • [Experimental Feature] Add the new feature for counting the peak inference GPU memory of the model.
• 2022-03-03: Add more methods, add the gpu warmup process in counting FPS and update the readme.
• 2021-09-29: Refactor again.
• 2021-08-31: Refactor.
• 2021-08-29: Create a new repository and upload the code.

@misc{MethodsCmp,
	author       = {Youwei Pang},
	title        = {MethodsCmp: A Simple Toolkit for Counting the FLOPs/MACs, Parameters and FPS of Pytorch-based Methods},
	howpublished = {\url{https://github.com/lartpang/MethodsCmp}},
	year         = {2021}
}

methodscmp's Issues

fvcore cannot count the FLOPs of the built-in module `nn.MultiHeadAttention` of PyTorch.