This tool is designed to compute the theoretical amount of multiply-add operations in neural networks. It can also compute the number of parameters and print per-layer computational cost of a given network.
ptflops
has two backends, pytorch
and aten
. pytorch
backend is a legacy one, it considers nn.Modules
only. However,
it's still useful, since it provides a better par-layer analytics for CNNs. In all other cases it's recommended to use
aten
backend, which considers aten operations, and therefore it covers more model architectures (including transformers).
The default backend is aten
. Please, don't use pytorch
backend for transformer architectures.
- aten.mm, aten.matmul, aten.addmm, aten.bmm
- aten.convolution
- Use
verbose=True
to see the operations which were not considered during complexity computation. - This backend prints per-module statistics only for modules directly nested into the root
nn.Module
. Deeper modules at the second level of nesting are not shown in the per-layer statistics. ignore_modules
option forcesptflops
to ignore the listed modules. This can be useful for research purposes. For instance, one can drop all convolutions from the counting process specifyingignore_modules=[torch.ops.aten.convolution, torch.ops.aten._convolution]
.
- Conv1d/2d/3d (including grouping)
- ConvTranspose1d/2d/3d (including grouping)
- BatchNorm1d/2d/3d, GroupNorm, InstanceNorm1d/2d/3d, LayerNorm
- Activations (ReLU, PReLU, ELU, ReLU6, LeakyReLU, GELU)
- Linear
- Upsample
- Poolings (AvgPool1d/2d/3d, MaxPool1d/2d/3d and adaptive ones)
Experimental support:
- RNN, LSTM, GRU (NLH layout is assumed)
- RNNCell, LSTMCell, GRUCell
- torch.nn.MultiheadAttention
- torchvision.ops.DeformConv2d
- visual transformers from timm
- This backend doesn't take into account some of the
torch.nn.functional.*
andtensor.*
operations. Therefore unsupported operations are not contributing to the final complexity estimation. Seeptflops/pytorch_ops.py:FUNCTIONAL_MAPPING,TENSOR_OPS_MAPPING
to check supported ops. Sometimes functional-level hooks conflict with hooks fornn.Module
(for instance, custom ones). In that case, counting with these ops can be disabled by passingbackend_specific_config={"count_functional" : False}
. ptflops
launches a given model on a random tensor and estimates amount of computations during inference. Complicated models can have several inputs, some of them could be optional. To construct non-trivial input one can use theinput_constructor
argument of theget_model_complexity_info
.input_constructor
is a function that takes the input spatial resolution as a tuple and returns a dict with named input arguments of the model. Next, this dict would be passed to the model as a keyword arguments.verbose
parameter allows to get information about modules that don't contribute to the final numbers.ignore_modules
option forcesptflops
to ignore the listed modules. This can be useful for research purposes. For instance, one can drop all convolutions from the counting process specifyingignore_modules=[torch.nn.Conv2d]
.
Pytorch >= 2.0. Use pip install ptflops==0.7.2.2
to work with torch 1.x.
From PyPI:
pip install ptflops
From this repository:
pip install --upgrade git+https://github.com/sovrasov/flops-counter.pytorch.git
import torchvision.models as models
import torch
from ptflops import get_model_complexity_info
with torch.cuda.device(0):
net = models.densenet161()
macs, params = get_model_complexity_info(net, (3, 224, 224), as_strings=True, backend='pytorch'
print_per_layer_stat=True, verbose=True)
print('{:<30} {:<8}'.format('Computational complexity: ', macs))
print('{:<30} {:<8}'.format('Number of parameters: ', params))
macs, params = get_model_complexity_info(net, (3, 224, 224), as_strings=True, backend='aten'
print_per_layer_stat=True, verbose=True)
print('{:<30} {:<8}'.format('Computational complexity: ', macs))
print('{:<30} {:<8}'.format('Number of parameters: ', params))
If ptflops was useful for your paper or tech report, please cite me:
@online{ptflops,
author = {Vladislav Sovrasov},
title = {ptflops: a flops counting tool for neural networks in pytorch framework},
year = 2018-2024,
url = {https://github.com/sovrasov/flops-counter.pytorch},
}
Thanks to @warmspringwinds and Horace He for the initial version of the script.
Model | Input Resolution | Params(M) | MACs(G) (pytorch ) |
MACs(G) (aten ) |
---|---|---|---|---|
alexnet | 224x224 | 61.10 | 0.72 | 0.71 |
convnext_base | 224x224 | 88.59 | 15.43 | 15.38 |
densenet121 | 224x224 | 7.98 | 2.90 | |
efficientnet_b0 | 224x224 | 5.29 | 0.41 | |
efficientnet_v2_m | 224x224 | 54.14 | 5.43 | |
googlenet | 224x224 | 13.00 | 1.51 | |
inception_v3 | 224x224 | 27.16 | 5.75 | 5.71 |
maxvit_t | 224x224 | 30.92 | 5.48 | |
mnasnet1_0 | 224x224 | 4.38 | 0.33 | |
mobilenet_v2 | 224x224 | 3.50 | 0.32 | |
mobilenet_v3_large | 224x224 | 5.48 | 0.23 | |
regnet_y_1_6gf | 224x224 | 11.20 | 1.65 | |
resnet18 | 224x224 | 11.69 | 1.83 | 1.81 |
resnet50 | 224x224 | 25.56 | 4.13 | 4.09 |
resnext50_32x4d | 224x224 | 25.03 | 4.29 | |
shufflenet_v2_x1_0 | 224x224 | 2.28 | 0.15 | |
squeezenet1_0 | 224x224 | 1.25 | 0.84 | 0.82 |
vgg16 | 224x224 | 138.36 | 15.52 | 15.48 |
vit_b_16 | 224x224 | 86.57 | 17.61 (wrong) | 16.86 |
wide_resnet50_2 | 224x224 | 68.88 | 11.45 |
Model | Input Resolution | Params(M) | MACs(G)