MutualNet: Adaptive ConvNet via Mutual Learning from Network Width and Resolution (ECCV'20 Oral) [arXiv]

This work proposes a method to train a network that is executable at dynamic resource constraints (e.g., FLOPs) during runtime. The proposed mutual learning scheme for input resolution and network width significantly improves the accuracy-efficiency tradeoffs over Slimmable Networks on various tasks such as image classification, object detection and instance segmentation. The method is also promising to serve as a plug-and-play strategy to boost a single network. It substantially outperforms a set of popular data augmentation and regularization methods. The overall framework is shown below.

We feed different sub-networks with different scaled images. This allows each sub-network to learn multi-scale representations from both network scale and input scale as illustrated below.

Install

PyTorch 1.0.1, torchvision 0.2.2, Numpy, pyyaml 5.1.
Follow the PyTorch example to prepare ImageNet dataset.

Run

Training

To train MobileNet v1, run the codes below

python train.py app:apps/mobilenet_v1.yml

Training hyperparameters are in the .yml files. width_mult_list is just used to print training logs for corresponding network widths. During testing, you can assign any desired width between the width lower bound and upper bound. To train other models, just use the corresponding .yml files.

Testing

Modify test_only: False to test_only: True in .yml file to enable testing.

Modify pretrained: /PATH/TO/YOUR/WEIGHTS to assign trained weights.

Modify width_mult_list to test more network widths.

python train.py app:apps/mobilenet_v1.yml

Note that this is to test the performance of all width-resolution configuration. You can obtain the query table by choosing the best-performed configuration at certain contraints (FLOPs). The query tables that we obtained for MobileNet v1 and MobileNet v2 are shown below.

MobileNet v1:

MFLOPs	569	518	466	418	380	342	309	269	239	215	187	166	146	127	111	76	63	41	30	21	13
Width-Reso	1.0-224	1.0-0.95	1.0-0.9	1.0-192	0.95-192	0.9-192	0.85-192	0.8-192	0.75-192	0.85-160	0.8-160	0.75-160	0.7-160	0.65-160	0.6-160	0.5-160	0.45-160	0.35-160	0.25-192	0.25-160	0.25-128
Acc	72.4	71.7	71.2	70.9	70.6	70.2	69.7	69.1	68.4	67.7	67.2	66.5	65.6	64.7	63.5	60.3	58.9	54.4	52.7	50.1	45.5

MobileNet v2:

MLOPs	300	284	269	254	221	209	198	187	164	154	128	107	89	84	57
Width-Reso	1.0-224	0.95-224	0.9-224	0.85-224	1.0-192	0.95-192	0.9-192	0.85-192	0.8-192	0.75-192	0.7-192	0.75-160	0.7-160	0.85-128	0.7-128
Acc	72.9	72.7	72.4	72.0	71.8	71.6	71.4	71.0	70.4	70.1	69.6	68.3	67.6	65.8	64.2

For deployment, you only need to keep the model weights and the query table, where the size of the query table is negligible. So the memory footprint (#params) is the same as a single model. If you want a more fine-grained query table, you can test with a smaller width stepsize.

Results and model weights

For those who do not have access to Google Drive: here is the link to all model weights in [BaiduYun]. The extraction code is 4y6m.

Performance over the whole FLOPs specturm

Comparison with US-Net under different backbones on ImageNet.

Model weights: [MobileNet v1], [MobileNet v2]

Scaling up model compared with EfficienNet

The best model scaling on MobileNet v1 compared with EfficientNet

Model	Best Model Scaling	FLOPs	Top-1 Acc
EfficientNet	d=1.4, w=1.2, r=1.3	2.3B	75.6%
MutualNet (Model)	w=1.6, r=1.3	2.3B	77.1%

Boosting performance of a single network

Top-1 accuracy on Cifar-10 and Cifar-100

WideResNet-28-10	GPU search hours	Cifar-10	Cifar-100
Baseline	0	96.1%	81.2%
Cutout	0	96.9%	81.6%
Mixup	0	97.3%	82.5%
AutoAugment	5000	97.4%	82.9%
Fast AutoAugment	3.5	97.3%	82.7%
MutualNet	0	97.3%	83.8%

Compared with state-of-the-art performance boosting methods on ImageNet

ResNet-50	Additional Cost	Top-1 Acc
Baseline	\	76.5%
Cutout	\	77.1%
Mixup	\	77.9%
CutMix	\	78.6%
KD	Teacher Network	76.5%
SENet	SE Block	77.6%
AutoAugment	15000 GPU search hours	77.6%
Fast AutoAugment	450 GPU search hours	77.6%
MutualNet (Model)	\	78.6%

Citation

If you find this useful in your work, please consider citing,

@inproceedings{yang2020mutualnet,
  title={MutualNet: Adaptive ConvNet via Mutual Learning from Network Width and Resolution},
  author={Yang, Taojiannan and Zhu, Sijie and Chen, Chen and Yan, Shen and Zhang, Mi and Willis, Andrew},
  booktitle={European Conference on Computer Vision},
  pages={299--315},
  year={2020},
  organization={Springer}
}

Reference

- The code is based on the implementation of Slimmable Networks.

gbyy422990/MutualNet