CondenseNet: Light weighted CNN for mobile devices
Clone or download
Pull request Compare This branch is 3 commits ahead, 3 commits behind ShichenLiu:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
models
.gitignore
LICENSE
README.md
__init__.py
layers.py
main.py
script.sh
utils.py

README.md

CondenseNets

This repository contains the code (in PyTorch) for "CondenseNet: An Efficient DenseNet using Learned Group Convolutions" paper by Gao Huang*, Shichen Liu*, Laurens van der Maaten and Kilian Weinberger (* Authors contributed equally).

Citation

If you find our project useful in your research, please consider citing:

@article{huang2017condensenet,
  title={CondenseNet: An Efficient DenseNet using Learned Group Convolutions},
  author={Huang, Gao and Liu, Shichen and van der Maaten, Laurens and Weinberger, Kilian Q},
  journal={arXiv preprint arXiv:1711.09224},
  year={2017}
}

Contents

  1. Introduction
  2. Usage
  3. Results
  4. Discussions
  5. Contacts

Introduction

CondenseNet is a novel, computationally efficient convolutional network architecture. It combines dense connectivity between layers with a mechanism to remove unused connections. The dense connectivity facilitates feature re-use in the network, whereas learned group convolutions remove connections between layers for which this feature re-use is superfluous. At test time, our model can be implemented using standard grouped convolutions —- allowing for efficient computation in practice. Our experiments demonstrate that CondenseNets are much more efficient than other compact convolutional networks such as MobileNets and ShuffleNets.

Figure 1: Learned Group Convolution with G=C=3.

Figure 2: CondenseNets with Fully Dense Connectivity and Increasing Growth Rate.

Usage

Dependencies

Train

As an example, use the following command to train a CondenseNet on ImageNet

python main.py --model condensenet -b 256 -j 20 /PATH/TO/IMAGENET \
--stages 4-6-8-10-8 --growth 8-16-32-64-128 --gpu 0,1,2,3,4,5,6,7 --resume

As another example, use the following command to train a CondenseNet on CIFAR-10

python main.py --model condensenet -b 64 -j 12 cifar10 \
--stages 14-14-14 --growth 8-16-32 --gpu 0 --resume

Evaluation

We take the ImageNet model trained above as an example.

To evaluate the trained model, use evaluate to evaluate from the default checkpoint directory:

python main.py --model condensenet -b 64 -j 20 /PATH/TO/IMAGENET \
--stages 4-6-8-10-8 --growth 8-16-32-64-128 --gpu 0 --resume \
--evaluate

or use evaluate-from to evaluate from an arbitrary path:

python main.py --model condensenet -b 64 -j 20 /PATH/TO/IMAGENET \
--stages 4-6-8-10-8 --growth 8-16-32-64-128 --gpu 0 --resume \
--evaluate-from /PATH/TO/BEST/MODEL

Note that these models are still the large models. To convert the model to group-convolution version as described in the paper, use the convert-from function:

python main.py --model condensenet -b 64 -j 20 /PATH/TO/IMAGENET \
--stages 4-6-8-10-8 --growth 8-16-32-64-128 --gpu 0 --resume \
--convert-from /PATH/TO/BEST/MODEL

Finally, to directly load from a converted model (that is, a CondenseNet), use a converted model file in combination with the evaluate-from option:

python main.py --model condensenet_converted -b 64 -j 20 /PATH/TO/IMAGENET \
--stages 4-6-8-10-8 --growth 8-16-32-64-128 --gpu 0 --resume \
--evaluate-from /PATH/TO/CONVERTED/MODEL

Other Options

We also include DenseNet implementation in this repository.
For more examples of usage, please refer to script.sh
For detailed options, please python main.py --help

Results

Results on ImageNet

Model FLOPs Params Top-1 Err. Top-5 Err. Pytorch Model
CondenseNet-74 (C=G=4) 529M 4.8M 26.2 8.3 Download (18.69M)
CondenseNet-74 (C=G=8) 274M 2.9M 29.0 10.0 Download (11.68M)

Results on CIFAR

Model FLOPs Params CIFAR-10 CIFAR-100
CondenseNet-50 28.6M 0.22M 6.22 -
CondenseNet-74 51.9M 0.41M 5.28 -
CondenseNet-86 65.8M 0.52M 5.06 23.64
CondenseNet-98 81.3M 0.65M 4.83 -
CondenseNet-110 98.2M 0.79M 4.63 -
CondenseNet-122 116.7M 0.95M 4.48 -
CondenseNet-182* 513M 4.2M 3.76 18.47

(* trained 600 epochs)

Inference time on ARM platform

Model FLOPs Top-1 Time(s)
VGG-16 15,300M 28.5 354
ResNet-18 1,818M 30.2 8.14
1.0 MobileNet-224 569M 29.4 1.96
CondenseNet-74 (C=G=4) 529M 26.2 1.89
CondenseNet-74 (C=G=8) 274M 29.0 0.99

Contact

liushichen95@gmail.com
gh349@cornell.com

We are working on the implementation on other frameworks.
Any discussions or concerns are welcomed!