This repository provides code for paper 《Lightweight Deep Neural Network Model With Padding-free Downsampling》

Framework & Performance

Experiment

I. For the CIFAR-100 dataset, using the first set of hyperparameters
I. For the CIFAR-100 dataset, using the second set of hyperparameters
II. For Stanford Dogs dataset
III. For ImageNet dataset
IV. Inference Latency
V. Ablation Experiments on CIFAR-100
VI. Comparison with other downsampling (EfficientFormerv2)
VII. For VegFru-292 dataset

Instructions for use

Getting started
Take MobileNetv3 as an example, when using our module
Acknowledgement

For the CIFAR-100 dataset, using the first set of hyperparameters

The first set of hyperparameters follows the settings of Haase et al.

《Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved MobileNets》
Daniel Haase∗   Manuel Amthor∗
ZEISS Microscopy   ZEISS Microscopy

Orig——CIFAR-100

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	3.066M	68.5M	75.37%
MobileNetv3-large(BSConv-S)	3.066M	68.5M	77.87%
ResNet-20	0.278M	41.4M	68.12%
ResNet-110(BSConv-U)	0.245M	41.8M	71.58%
WideResNet-40-3	5.056M	735.8M	76.23%
WideResNet-40-8(BSConv-U)	4.286M	675.1M	77.79%

Ours——CIFAR-100

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	3.067M	54.6M ↓	75.71%
MobileNetv3-large(BSConv-S)	3.067M	54.6M ↓	78.36%
ResNet-20	0.282M	37.8M ↓	68.30%
ResNet-110(BSConv-U)	0.249M	38.6M ↓	71.62%
WideResNet-40-3	5.287M	668.7M ↓	76.28%
WideResNet-40-8(BSConv-U)	4.457M	615.6M ↓	78.05%

For the CIFAR-100 dataset, using the second set of hyperparameters

Orig——CIFAR-100

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	4.330M	68.8M	76.00%
Parc-MobileNet-v2	2.348M	91.3M	76.20%
GhostNet	4.029M	44.6M	74.00%
ShuffleNet-v2	1.356M	46.2M	70.90%

Ours——CIFAR-100

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	4.331M	54.7M ↓	76.60%
Parc-MobileNet-v2	2.348M	73.0M ↓	76.60%
GhostNet	4.030M	34.8M ↓	74.10%
ShuffleNet-v2	1.358M	35.7M ↓	71.50%

For Stanford Dogs dataset

Orig——Stanford Dogs

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	3.086M	230.1M	51.07%
MobileNetv3-large-bsconvs	3.086M	230.1M	59.68%

Ours——Stanford Dogs

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	3.087M	212.6M ↓	54.11%
MobileNetv3-large-bsconvs	3.087M	212.6M ↓	60.79%

For ImageNet dataset

Orig——ImageNet

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	5.480M	232.5M	69.50%

Ours——ImageNet

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	5.481M	214.9M ↓	69.50%

Inference Latency

Orig——Latency

Model	AMD Ryzen 5 5600H	MediaTek Tiangui 1000+
MobileNetv3-large	8.5ms	27.0ms
Parc-MobileNet-v2	8.7ms	37.4ms
GhostNet	11.4ms	36.6ms
ShuffleNet-v2	6.2ms	19.4ms

Ours——Latency

Model	AMD Ryzen 5 5600H	MediaTek Tiangui 1000+
MobileNetv3-large	9.0ms	26.3ms ↓
Parc-MobileNet-v2	9.3ms	34.0ms ↓
GhostNet	11.7ms	26.8ms ↓
ShuffleNet-v2	7.4ms	18.8ms ↓

Ablation Experiments on CIFAR-100

Model	orig	+Stem	+Downsampling	ours
MobileNetv3-large	76.0%	75.9%	76.4%	76.6%↑
Parc-MobileNet-v2	76.2%	76.6%	76.4%	76.6%↑
GhostNet	76.0%	74.2%	73.8%	74.1%↑
ShuffleNet-v2	70.9%	72.0%	70.4%	71.5%↑

Comparison with other downsampling (EfficientFormerv2)

《Rethinking Vision Transformers for MobileNet Size and Speed》
Yanyu Li
Snap Inc. Northeastern University

EfficientFormerv2-Downsampling——CIFAR-100

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	4.317M	78.0M	75.80%
Parc-MobileNet-v2	2.558M	97.5M	75.70%
GhostNet	4.092M	58.3M	74.30%
ShuffleNet-v2	2.804M	84.1M	70.60%

Ours——CIFAR-100

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	4.331M	54.7M	76.60%
Parc-MobileNet-v2	2.348M	73.0M	76.60%
GhostNet	4.030M	34.8M	74.10%
ShuffleNet-v2	1.358M	35.7M	71.50%

For VegFru-292 dataset

Orig——VegFru-292

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	4.576M	224.5M	89.20%
Parc-MobileNet-v2	2.605M	314.8M	89.10%
GhostNet	4.276M	147.9M	89.60%
ShuffleNet-v2	1.553M	148.1M	88.40%

Ours——VegFru-292

Model	Parameters	FLOPs	Accuracy
MobileNetv3-large	4.577M	205.7M ↓	89.90%
Parc-MobileNet-v2	2.605M	305.5M ↓	90.00%
GhostNet	4.276M	136.9M ↓	90.30%
ShuffleNet-v2	1.554M	130.7M ↓	87.70%

Getting started

• For the BSConv folder, when using it for the first time, use "--download" to download the dataset.

python bsconv_pytorch_train.py --data-root cifar100 --dataset cifar100 --architecture cifar_mobilenetv3_large_w1 --download --gpu-id 0

• "--data-root" is the dataset path, "--dataset" is the dataset name, "--architecture" is the model name.

python bsconv_pytorch_train.py --data-root cifar100 --dataset cifar100 --architecture cifar_mobilenetv3_large_w1 --gpu-id 0
python bsconv_pytorch_train.py --data-root cifar100 --dataset cifar100 --architecture cifar_mobilenetv3_large_w1_bsconvs_p1d6 --gpu-id 0
python bsconv_pytorch_train.py --data-root cifar100 --dataset cifar100 --architecture cifar_wrn40_3 --gpu-id 0
python bsconv_pytorch_train.py --data-root cifar100 --dataset cifar100 --architecture cifar_wrn40_8_bsconvu --gpu-id 0
python bsconv_pytorch_train.py --data-root cifar100 --dataset cifar100 --architecture cifar_resnet20 --gpu-id 0
python bsconv_pytorch_train.py --data-root cifar100 --dataset cifar100 --architecture cifar_resnet110_bsconvu --gpu-id 0 

Take MobileNetv3 as an example, when using our module.

• Replace init_conv on line 321 in the mobilenet.py file with our stem layer.

self.backbone.add_module("init_conv", StemBlock(in_channels, init_conv_channels))

• Uncomment the if stride==2 on lines 157, 168, 237, and 261 in the common.py file.

        if stride == 2:
          self.maxx = nn.MaxPool2d(kernel_size=3, stride=2,padding=0)

        if self.stride == 2:
          b = self.maxx(b)
          return x + b

    if stride ==2:
          return ConvBlock(
             in_channels=channels,
             out_channels=channels,
             kernel_size=3,
             stride=stride,
             padding=0,
             groups=channels,
             use_bn=use_bn,
             activation=activation)

    if stride ==2 :
        return ConvBlock(
             in_channels=channels,
             out_channels=channels,
             kernel_size=5,
             stride=stride,
             padding=1,
             groups=channels,
             use_bn=use_bn,
             activation=activation)

Acknowledgement

Thanks to BSConv and EfficientFormerv2. Our code is based on the BSConv library and EfficientFormerv2 library.