https://github.com/lartpang/methodscmp

MethodsCmp: A Simple Toolkit for Counting the FLOPs/MACs, Parameters and FPS of Pytorch-based Methods
https://github.com/lartpang/methodscmp

flops flops-counter fps fps-counter macs macs-counter memory memory-counter parameter-counter parameters python pytorch

Last synced: 26 days ago
JSON representation

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

• Host: GitHub
• URL: https://github.com/lartpang/methodscmp
• Owner: lartpang
• Created: 2021-08-29T04:56:20.000Z (over 3 years ago)
• Default Branch: main
• Last Pushed: 2024-05-13T09:06:43.000Z (8 months ago)
• Last Synced: 2024-12-01T19:51:36.389Z (about 1 month ago)
• Topics: flops, flops-counter, fps, fps-counter, macs, macs-counter, memory, memory-counter, parameter-counter, parameters, python, pytorch
• Language: Python
• Homepage:
• Size: 106 KB
• Stars: 18
• Watchers: 4
• Forks: 4
• Open Issues: 0
• Metadata Files:
  • Readme: readme.md

Awesome Lists containing this project

README

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

```shell

$ 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](https://github.com/lartpang/MethodsCmp/issues/1)

## Usage

```bash

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:

```shell

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.

```latex

@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}

}

```