FlopCo is a Python library that aims to make FLOPs and MACs counting simple and accessible for Pytorch neural networks. Moreover FlopCo allows to collect other useful model statistics, such as number of parameters, shapes of layer inputs/outputs, etc.
- numpy
- pytorch
pip install flopco-pytorch
from flopco import FlopCo
from torchvision.models import resnet50
device = 'cuda'
model = resnet50().to(device)
# Estimate model statistics by making one forward pass througth the model,
# for the input image of size 3 x 224 x 224
stats = FlopCo(model, img_size = (1, 3, 224, 224), device = device)
print(stats.total_macs, stats.relative_flops)
List of estimated statistics includes:
- total number of FLOPs/MACs/parameters
- number of FLOPs/MACs/parameters for each layer
- relative number of FLOPs/MACs/parameters for each layer
- input/output shapes for each layer
By default for statistics counting nn.Conv2d and nn.Linear layers are used.
To include more layer types in computation, pass instances
to the constructor
stats = FlopCo(model,
img_size = (1, 3, 224, 224),
device = device,
instances = [nn.Conv2d, nn.Linear,\
nn.BatchNorm2d, nn.ReLU,\
nn.MaxPool2d, nn.AvgPool2d,\
nn.Softmax]
)
Project is distributed under MIT License