PydMobileNet: Deep Residual Networks with Pyramid Depthwise Separable Convolution

Dependencies

1. Mxnet >= 1.3.0. To install the mxnet, follow the instruction in https://mxnet.incubator.apache.org/install/index.html
2. Numpy. By command: 'pip install numpy'
3. Gluoncv. By command: 'pip install gluoncv'

Main results

Network architecture (CIFAR)

Group	Output size	Networks
Image	32x32x3
Convolution	32x32x32	3x3 Convolution
Stage 1	32x32x32	Residual block x6
Stage 2	16x16x64	Residual block x6
Stage 3	8x8x128	Residual block x6
Pooling	1x1x128	8x8 Global Average Pooling
Classifier	1x1x10/100	10/100D Fully-Connected

CIFAR

• Number of channels: {32, 32, 64, 128, 128}

Model	Depth	#Params	FLOPs	CIFAR-10	CIFAR-100
ResNet-29-0.5	29	0.221M	29M	6.97	19.62
MobileNet-29-0.5	29	0.079M	12M	8.63	22.59
MobileNet-29-1	29	0.142M	22M	7.09	19.40
MobileNet-29-1.5	29	0.206M	32M	6.56	18.09
PydMobileNet-Add-29-0.25	29	0.060M	10M	9.43	21.96
PydMobileNet-Add-29-0.5	29	0.104M	18M	7.29	20.26
PydMobileNet-Add-29-0.75	29	0.148M	26M	6.52	17.95
PydMobileNet-Add-29-1	29	0.193M	34M	6.00	17.54
PydMobileNet-Concat-29-0.25	29	0.092M	14M	7.33	21.04
PydMobileNet-Concat-29-0.5	29	0.170M	27M	5.71	17.27
PydMobileNet-Concat-29-0.75	29	0.247M	39M	5.68	16.28
ResNet-56-0.5	56	0.435M	60M	5.76	17.60
MobileNet-56-0.5	56	0.151M	23M	6.75	18.56
MobileNet-56-1	56	0.283M	43M	6.02	17.15
MobileNet-56-1.5	56	0.416M	63M	5.29	16.58
PydMobileNet-Add-56-0.25	56	0.109M	19M	7.38	20.41
PydMobileNet-Add-56-0.5	56	0.200M	36M	6.19	17.36
PydMobileNet-Add-56-0.75	56	0.292M	52M	5.55	16.58
PydMobileNet-Add-56-1	56	0.382M	69M	4.98	16.23
PydMobileNet-Concat-56-0.25	56	0.175M	28M	6.23	17.85
PydMobileNet-Concat-56-0.5	56	0.332M	53M	5.24	15.67
PydMobileNet-Concat-56-0.75	56	0.489M	79M	4.72	14.60

Notes: #Params are for CIFAR-10 dataset. For CIFAR-100: #Params + (128 + 1) * 100 - (128 + 1) * 10

ImageNet32

• Number of channels: {64, 64, 128, 256, 1280}

Model	Depth	#Params	Top-1	Top-5
WRN-28-2	28	1.6M	56.92	30.92
WRN-28-5	28	9.5M	45.36	21.36
WRN-28-10	28	37.1M	40.96	18.87
PydMobileNet-Concat-56-0.25	56	2.23M	48.18	23.70
PydMobileNet-Concat-56-0.5	56	2.80M	44.70	20.58
PydMobileNet-Concat-56-0.75	56	3.37M	41.99	18.59

Run validation

• To validate all models, run 'python run_validation.py'
• To see all configuration, run 'python run_validation.py --help'
• Note: program needs internet connection to download the CIFAR-10 and CIFAR-100 dataset from mxnet website at the first running.

Citation

Please cite these papers in your publications if it helps your research:

@ARTICLE{thanh2018arxiv,
  author    = {Van{-}Thanh, Hoang and Kang{-}Hyun, Jo},
  title     = {PydMobileNet: Improved Version of MobileNets with Pyramid Depthwise Separable Convolution},
  journal   = {CoRR},
  volume    = {abs/1811.07083},
  year      = {2018},
  url       = {http://arxiv.org/abs/1811.07083},
  archivePrefix = {arXiv},
  eprint    = {1811.07083},
}