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}
}

Introduction
Usage
Results
Discussions
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!

deepxkn/CondenseNet