swagshaw/mindspore-flops

Count the MACs / FLOPs of your MindSpore model.

msflops

This script is designed to compute the theoretical amount of multiply-add operations in convolutional neural networks. It also can compute the number of parameters(TODO) and print per-layer computational cost of a given network.

Note: most of the code I have given to the mindspore community, the subsequent may be integrated into the mindspore profile function.

Supported layers:

• Conv1d/2d/3d (including grouping)
• BatchNorm1d/2d/3d, GroupNorm, InstanceNorm1d/2d/3d
• Activations (ReLU, PReLU, ELU, ReLU6, LeakyReLU)
• Dense
• Poolings (AvgPool1d/2d/3d, MaxPool1d/2d/3d and adaptive ones)

Features & TODO

Note: These features work only MindSpore.nn. Modules in MindSpore.ops are not supported yet.

• FLOPs
• Number of Parameters
• Total memory
• Madd(FMA)
• MemRead
• MemWrite
• Model summary(detail, layer-wise)
• Export score table
• Arbitrary input shape

Example

import sys
import mindspore
from msflops import stat
from models import alexnet
from mindspore import context

if __name__ == '__main__':
    model = alexnet.AlexNet() # which is a model designed by yourself
    context.set_context(mode=context.PYNATIVE_MODE)
    stat(model, (3, 224, 224))

Benchmark

TODO:Fill a table about the data of more models in mindspore.modelzoo. And fix some errors between pytorch and mindspore results.

Model	Input Resolution	Flops
alexnet	224x224	710.6(M)
resnet18	224x224	1.82(G)
resnet34	224x224	3.67(G)
resnet50	224x224	4.11(G)
resnet101	224x224	7.83(G)
resnet152	224x224	11.55(G)
vgg16	224x224	15.49(G)
vgg16_bn	224x224	15.51(G)
vgg11	224x224	7.62(G)
vgg11_bn	224x224	7.63(G)
vgg13	224x224	11.32(G)
vgg13_bn	224x224	11.34(G)
vgg19	224x224	19.65(G)
vgg19_bn	224x224	19.66(G)
densenet121	224x224	2.87(G)
densenet161	224x224	7.79(G)
densenet129	224x224	3.4(G)
densenet201	224x224	4.34(G)
squeezenet	224x224	715(M)
squeezenet_res	224x224	711.5(M)
shufflenetv1	224x224	536.78(M)
shufflenetv2	224x224	149.6(M)
efficientnetb0	224x224	392.55(M)

Toread

There are still a lot of errors and imperfections, so I did not package to pypi. most of the code I have given to the mindspore community, the subsequent may be integrated into the mindspore profile function. This repository can be used as a lightweight tool alone, I will continue to improve it when I have time, and I welcome everyone to contribute and merge code.

References

Thanks to @Swall0w & @Lyken17 for the initial version of flops computation, @Swall0w for the backbone of scripts.

• pytorch-OpCounter
• torchstat.
• flops-counter.pytorch
• pytorch_model_summary
• chainer_computational_cost
• convnet-burden.