Large-scale image classification networks for embedded systems

This repository contains several classification models on MXNet/Gluon and PyTorch and scripts for trainig/converting models. All models are designed for using with ImageNet-1k dataset.

Requirements

All models/scripts are on Python. Tested on Python 2.7 and 3.6.

For Gluon way

If you only want to use models:

mxnet >= 1.2.1

If you want also training models:

gluoncv >= 0.2.0

For PyTorch way

torch >= 0.4.1
torchvision >= 0.2.1

List of models

• ResNet ('Deep Residual Learning for Image Recognition')
• PreResNet ('Identity Mappings in Deep Residual Networks')
• DenseNet ('Densely Connected Convolutional Networks')
• CondenseNet ('Condense````Net: An Efficient DenseNet using Learned Group Convolutions')
• DarkNet ('Darknet: Open source neural networks in c')
• SqueezeNet ('SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size')
• SqueezeNext ('SqueezeNext: Hardware-Aware Neural Network Design')
• ShuffleNet ('ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices')
• ShuffleNetV2 ('ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design')
• MENet ('Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications')
• MobileNet ('MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications')
• FD-MobileNet ('FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy')
• MobileNetV2 ('MobileNetV2: Inverted Residuals and Linear Bottlenecks')

Pretrained models

Some remarks:

• All pretrained models can be downloaded automatically during use (use the parameter pretrained).
• Top1/Top5 are the standard 1-crop Top-1/Top-5 errors (in percents) on the validation subset of the ImageNet1k dataset.
• ResNet/PreResNet with b-suffix is a version of the networks with the stride in the second convolution of the bottleneck block. Respectively a network without b-suffix has the stride in the first convolution.
• ResNet/PreResNet models do not use biasses in convolutions at all.
• CondenseNet models are only so-called converted versions.

For Gluon

Model	Top1	Top5	Params	FLOPs	Remarks
ResNet-10	37.09	15.55	5,418,792	892.62M	Training (log)
ResNet-12	37.62	15.56	5,492,776	1,124.23M	1-stage training (log)
ResNet-14	36.17	14.52	5,788,200	1,355.64M	1-stage training (log)
ResNet-16	33.57	12.50	6,968,872	1,586.95M	1-stage training (log)
ResNet-18 x0.25	53.13	27.77	831,096	136.64M	1-stage training (log)
ResNet-18 x0.5	38.94	16.46	3,055,880	485.22M	1-stage training (log)
ResNet-18	29.13	9.94	11,689,512	1,818.21M	Training (log)
ResNet-34	25.34	7.92	21,797,672	3,669.16M	Converted from Gluon Model Zoo (log)
ResNet-50	23.50	6.87	25,557,032	3,868.96M	Converted from Gluon Model Zoo (log)
ResNet-50b	22.92	6.44	25,557,032	4,100.70M	Converted from Gluon Model Zoo (log)
ResNet-101	21.66	5.99	44,549,160	7,586.30M	Converted from Gluon Model Zoo (log)
ResNet-101b	21.18	5.60	44,549,160	7,818.04M	Converted from Gluon Model Zoo (log)
ResNet-152	21.01	5.61	60,192,808	11,304.85M	Converted from Gluon Model Zoo (log)
ResNet-152b	20.54	5.37	60,192,808	11,536.58M	Converted from Gluon Model Zoo (log)
PreResNet-18	29.45	10.29	11,687,848	1,818.41M	Converted from Gluon Model Zoo (log)
PreResNet-34	25.88	8.11	21,796,008	3,669.36M	Converted from Gluon Model Zoo (log)
PreResNet-50	23.39	6.68	25,549,480	3,869.16M	Converted from Gluon Model Zoo (log)
PreResNet-50b	23.16	6.64	25,549,480	4,100.90M	Converted from Gluon Model Zoo (log)
PreResNet-101	21.45	5.75	44,541,608	7,586.50M	Converted from Gluon Model Zoo (log)
PreResNet-101b	21.73	5.88	44,541,608	7,818.24M	Converted from Gluon Model Zoo (log)
PreResNet-152	23.00	6.34	60,185,256	11,305.05M	Converted from Gluon Model Zoo (log)
PreResNet-152b	21.00	5.75	60,185,256	11,536.78M	Converted from Gluon Model Zoo (log)
DenseNet-121	25.11	7.80	7,978,856	2,852.39M	Converted from Gluon Model Zoo (log)
DenseNet-161	22.40	6.18	28,681,000	7,761.25M	Converted from Gluon Model Zoo (log)
DenseNet-169	23.89	6.89	14,149,480	3,381.48M	Converted from Gluon Model Zoo (log)
DenseNet-201	22.71	6.36	20,013,928	4,318.75M	Converted from Gluon Model Zoo (log)
CondenseNet-74 (C=G=4)	26.82	8.64	4,773,944	533.64M	From ShichenLiu/CondenseNet (log)
CondenseNet-74 (C=G=8)	29.76	10.49	2,935,416	278.55M	From ShichenLiu/CondenseNet (log)
SqueezeNet v1.0	42.81	19.98	1,248,424	828.30M	Converted from Gluon Model Zoo (log)
SqueezeNet v1.1	43.06	20.23	1,235,496	354.88M	Converted from Gluon Model Zoo (log)
108-MENet-8x1 (g=3)	46.11	22.37	654,516	40.64M	From clavichord93/MENet (log)
128-MENet-8x1 (g=4)	45.80	21.93	750,796	43.58M	From clavichord93/MENet (log)
228-MENet-12x1 (g=3)	35.03	13.99	1,806,568	148.93M	From clavichord93/MENet (log)
256-MENet-12x1 (g=4)	34.49	13.90	1,888,240	146.11M	From clavichord93/MENet (log)
348-MENet-12x1 (g=3)	31.17	11.41	3,368,128	306.31M	From clavichord93/MENet (log)
352-MENet-12x1 (g=8)	34.70	13.75	2,272,872	151.03M	From clavichord93/MENet (log)
456-MENet-24x1 (g=3)	29.57	10.43	5,304,784	560.72M	From clavichord93/MENet (log)
MobileNet x0.25	48.37	24.10	470,072	42.30M	Converted from Gluon Model Zoo (log)
MobileNet x0.5	37.37	15.37	1,331,592	152.04M	Converted from Gluon Model Zoo (log)
MobileNet x0.75	32.71	12.28	2,585,560	329.22M	Converted from Gluon Model Zoo (log)
MobileNet x1.0	29.25	10.03	4,231,976	573.83M	Converted from Gluon Model Zoo (log)
FD-MobileNet x0.25	56.73	31.99	383,160	12.44M	From clavichord93/FD-MobileNet (log)
FD-MobileNet x0.5	44.66	21.08	993,928	40.93M	From clavichord93/FD-MobileNet (log)
FD-MobileNet x1.0	35.95	14.72	2,901,288	146.08M	From clavichord93/FD-MobileNet (log)
MobileNetV2 x0.25	48.89	25.24	1,516,392	32.22M	Converted from Gluon Model Zoo (log)
MobileNetV2 x0.5	35.51	14.64	1,964,736	95.62M	Converted from Gluon Model Zoo (log)
MobileNetV2 x0.75	30.82	11.26	2,627,592	191.61M	Converted from Gluon Model Zoo (log)
MobileNetV2 x1.0	28.51	9.90	3,504,960	320.19M	Converted from Gluon Model Zoo (log)

For PyTorch

Model	Top1	Top5	Params	FLOPs	Remarks
ResNet-10	37.46	15.85	5,418,792	892.62M	Converted from GL model (log)
ResNet-12	38.02	15.89	5,492,776	1,124.23M	Converted from GL model (log)
ResNet-14	36.50	14.84	5,492,776	1,355.64M	Converted from GL model (log)
ResNet-16	33.73	12.87	6,968,872	1,586.95M	Converted from GL model (log)
ResNet-18 x0.25	53.37	28.06	831,096	136.64M	Converted from GL model (log)
ResNet-18 x0.5	39.31	16.79	3,055,880	485.22M	Converted from GL model (log)
ResNet-18	29.52	10.21	11,689,512	1,818.21M	Converted from GL model (log)
ResNet-34	25.66	8.18	21,797,672	3,669.16M	Converted from Gluon Model Zoo (log)
ResNet-50	23.79	7.05	25,557,032	3,868.96M	Converted from Gluon Model Zoo (log)
ResNet-50b	23.05	6.65	25,557,032	4,100.70M	Converted from Gluon Model Zoo (log)
ResNet-101	21.90	6.22	44,549,160	7,586.30M	Converted from Gluon Model Zoo (log)
ResNet-101b	21.45	5.81	44,549,160	7,818.04M	Converted from Gluon Model Zoo (log)
ResNet-152	21.26	5.82	60,192,808	11,304.85M	Converted from Gluon Model Zoo (log)
ResNet-152b	20.74	5.50	60,192,808	11,536.58M	Converted from Gluon Model Zoo (log)
PreResNet-18	29.76	10.57	11,687,848	1,818.41M	Converted from Gluon Model Zoo (log)
PreResNet-34	26.23	8.41	21,796,008	3,669.36M	Converted from Gluon Model Zoo (log)
PreResNet-50	23.70	6.85	25,549,480	3,869.16M	Converted from Gluon Model Zoo (log)
PreResNet-50b	23.33	6.87	25,549,480	4,100.90M	Converted from Gluon Model Zoo (log)
PreResNet-101	21.74	5.91	44,541,608	7,586.50M	Converted from Gluon Model Zoo (log)
PreResNet-101b	21.95	6.03	44,541,608	7,818.24M	Converted from Gluon Model Zoo (log)
PreResNet-152	23.20	6.53	60,185,256	11,305.05M	Converted from Gluon Model Zoo (log)
PreResNet-152b	21.34	5.91	60,185,256	11,536.78M	Converted from Gluon Model Zoo (log)
DenseNet-121	25.57	8.03	7,978,856	2,852.39M	Converted from Gluon Model Zoo (log)
DenseNet-161	22.86	6.44	28,681,000	7,761.25M	Converted from Gluon Model Zoo (log)
DenseNet-169	24.40	7.19	14,149,480	3,381.48M	Converted from Gluon Model Zoo (log)
DenseNet-201	23,10	6.63	20,013,928	4,318.75M	Converted from Gluon Model Zoo (log)
CondenseNet-74 (C=G=4)	26.25	8.28	4,773,944	533.64M	From ShichenLiu/CondenseNet (log)
CondenseNet-74 (C=G=8)	28.93	10.06	2,935,416	278.55M	From ShichenLiu/CondenseNet (log)
SqueezeNet v1.0	41.91	19.58	1,248,424	828.30M	Converted from TorchVision (log)
SqueezeNet v1.1	41.82	19.38	1,235,496	354.88M	Converted from TorchVision (log)
108-MENet-8x1 (g=3)	43.92	20.76	654,516	40.64M	From clavichord93/MENet (log)
128-MENet-8x1 (g=4)	43.95	20.62	750,796	43.58M	From clavichord93/MENet (log)
228-MENet-12x1 (g=3)	33.57	13.28	1,806,568	148.93M	From clavichord93/MENet (log)
256-MENet-12x1 (g=4)	33.41	13.26	1,888,240	146.11M	From clavichord93/MENet (log)
348-MENet-12x1 (g=3)	30.10	10.92	3,368,128	306.31M	From clavichord93/MENet (log)
352-MENet-12x1 (g=8)	33.31	13.08	2,272,872	151.03M	From clavichord93/MENet (log)
456-MENet-24x1 (g=3)	28.40	9.93	5,304,784	560.72M	From clavichord93/MENet (log)
MobileNet x0.25	49.13	24.93	470,072	42.30M	Converted from Gluon Model Zoo (log)
MobileNet x0.5	38.12	15.99	1,331,592	152.04M	Converted from Gluon Model Zoo (log)
MobileNet x0.75	33.54	12.85	2,585,560	329.22M	Converted from Gluon Model Zoo (log)
MobileNet x1.0	29.86	10.36	4,231,976	573.83M	Converted from Gluon Model Zoo (log)
FD-MobileNet x0.25	55.77	31.32	383,160	12.44M	From clavichord93/FD-MobileNet (log)
FD-MobileNet x0.5	43.85	20.72	993,928	40.93M	From clavichord93/FD-MobileNet (log)
FD-MobileNet x1.0	34.70	14.05	2,901,288	146.08M	From clavichord93/FD-MobileNet (log)
MobileNetV2 x0.25	49.72	25.87	1,516,392	32.22M	Converted from Gluon Model Zoo (log)
MobileNetV2 x0.5	36.54	15.19	1,964,736	95.62M	Converted from Gluon Model Zoo (log)
MobileNetV2 x0.75	31.89	11.76	2,627,592	191.61M	Converted from Gluon Model Zoo (log)
MobileNetV2 x1.0	29.31	10.39	3,504,960	320.19M	Converted from Gluon Model Zoo (log)