S²-BNN (Self-supervised Binary Neural Networks Using Distillation Loss)

This is the official pytorch implementation of our paper:

"S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-bit Neural Networks via Guided Distribution Calibration" (CVPR 2021)

by Zhiqiang Shen, Zechun Liu, Jie Qin, Lei Huang, Kwang-Ting Cheng and Marios Savvides.

In this paper, we introduce a simple yet effective self-supervised approach using distillation loss for learning efficient binary neural networks. Our proposed method can outperform the simple contrastive learning baseline (MoCo V2) by an absolute gain of 5.5∼15% on ImageNet.

The student models are not restricted to the binary neural networks, you can replace with any efficient/compact models.

News

[Aug. 18, 2021] Add ResNet-50 result as the student (real-valued model) with SwAV/RN50-w4 as the teacher.

Models	Pre-train epochs	batch size	linear cls. log	Top-1 (%)	Pre-train models
SimSiam	100	512	log	68.1	GitHub
S2-BNN (Distillation_only)	100	512	log	68.7	Download

Note that we use the same linear evaluation procedure as SimSiam.

Citation

If you find our code is helpful for your research, please cite:

@InProceedings{Shen_2021_CVPR,
	author    = {Shen, Zhiqiang and Liu, Zechun and Qin, Jie and Huang, Lei and Cheng, Kwang-Ting and Savvides, Marios},
	title     = {S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-Bit Neural Networks via Guided Distribution Calibration},
	booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
	year      = {2021}

}

Preparation

1. Requirements:

• Python
• PyTorch
• Torchvision

2. Data:

• Download ImageNet dataset following https://github.com/pytorch/examples/tree/master/imagenet#requirements.

Training & Testing

To train a model, run the following scripts. All our models are trained with 8 GPUs.

1. Standard Two-Step Training:

Our enhanced MoCo V2:

Step 1:

cd Contrastive_only/step1
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders]  --mlp --moco-t 0.2 --aug-plus --cos -j 48  

Step 2:

cd Contrastive_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders]  --mlp --moco-t 0.2 --aug-plus --cos -j 48  --model-path ../step1/checkpoint_0199.pth.tar

Our MoCo V2 + Distillation Loss:

Download real-valued teacher network here. We use MoCo V2 800-epoch pretrained model, while you can choose other stronger self-supervised models as the teachers.

Step 1:

cd Contrastive+Distillation/step1
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0  --teacher-path ../../moco_v2_800ep_pretrain.pth.tar 

Step 2:

cd Contrastive+Distillation/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0  --teacher-path ../../moco_v2_800ep_pretrain.pth.tar --model-path ../step1/checkpoint_0199.pth.tar

Our Distillation Loss Only:

Step 1:

cd Distillation_only/step1
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar 

Step 2:

cd Distillation_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar --model-path ../step1/checkpoint_0199.pth.tar

2. Simple One-Step Training (Conventional):

Our enhanced MoCo V2:

cd Contrastive_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 

Our MoCo V2 + Distillation Loss:

cd Contrastive+Distillation/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar 

Our Distillation Loss Only:

cd Distillation_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar 

You can replace binary neural networks with any kinds of efficient/compact models on one-step training.

3. Testing:

• To linearly evaluate a model, run the following script:

  python main_lincls.py  --lr 0.1  -j 24  --batch-size 256  --pretrained  /home/szq/projects/s2bnn/checkpoint_0199.pth.tar --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] 
  

Results & Models

We provide pre-trained models with different training strategies, we report in the table #Epochs, FLOPs, OPs, Top-1 accuracy on ImageNet validation set:

Models	#Epoch	FLOPs (x108)	OPs (x108)	Top-1 (%)	Trained models
MoCo V2 baseline	200	0.12	0.87	46.9	Download
Our enhanced MoCo V2	200	0.12	0.87	52.5	Download
Our MoCo V2 + Distillation Loss	200	0.12	0.87	56.0	Download
Our Distillation Loss Only	200	0.12	0.87	61.5	Download

Training Logs

Our linear evaluation logs are availabe at here.

Acknowledgement

MoCo V2 (Improved Baselines with Momentum Contrastive Learning)

ReActNet (ReActNet: Towards Precise Binary NeuralNetwork with Generalized Activation Functions)

MEAL V2 (MEAL V2: Boosting Vanilla ResNet-50 to 80%+ Top-1 Accuracy on ImageNet without Tricks)

Contact

Zhiqiang Shen, CMU (zhiqiangshen0214 at gmail.com)