Pre-trained models

This repository contains pretrained models. (converted from gluon-cv)

Environment

• PyTorch 1.1
• Python 3.6
• OpenCV

Evaluation on imagenet

resnet

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
ResNet18_v1	70.93	89.92	70.18	89.52
ResNet34_v1	74.37	91.87	74.04	91.82
ResNet50_v1	77.36	93.57	77.16	93.56
ResNet101_v1	78.34	94.01	78.23	94.09
ResNet152_v1	79.22	94.64
ResNet18_v2	71.00	89.92	70.10	89.48
ResNet34_v2	74.40	92.08	74.37	92.02
ResNet50_v2	77.11	93.43	77.00	93.36
ResNet101_v2	78.53	94.17	78.52	94.15
ResNet152_v2	79.21	94.31

resnet_v1b

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
ResNet18_v1b	70.94	89.83	70.08	89.44
ResNet34_v1b	74.65	92.08	74.11	92.16
ResNet50_v1b	77.67	93.82	77.57	93.58
ResNet50_v1b_gn	77.36	93.59	77.22	93.54
ResNet101_v1b	79.20	94.61	79.12	94.47
ResNet152_v1b	79.69	94.74	78.07	93.97
ResNet50_v1c	78.03	94.09	77.89	94.02
ResNet101_v1c	79.60	94.75	79.48	94.72
ResNet152_v1c	80.01	94.96	78.18	93.99
ResNet50_v1d	79.15	94.58	79.04	94.61
ResNet101_v1d	80.51	95.12	80.52	95.23
ResNet152_v1d	80.61	95.34	80.75	95.34

• ResNet_v1b modifies ResNet_v1 by setting stride at the 3x3 layer for a bottleneck block.
• ResNet_v1c modifies ResNet_v1b by replacing the 7x7 conv layer with three 3x3 conv layers.
• ResNet_v1d modifies ResNet_v1c by adding an avgpool layer 2x2 with stride 2 downsample feature map on the residual path to preserve more information.

mobilenet

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
MobileNet1.0	73.28	91.30	72.85	91.12
MobileNet0.75	70.25	89.49	69.85	89.46
MobileNet0.5	65.20	86.34	64.19	85.71
MobileNet0.25	52.91	76.94	51.09	75.36
MobileNetV2_1.0	71.92	90.56	71.78	90.36
MobileNetV2_0.75	69.61	88.95	69.29	88.81
MobileNetV2_0.5	64.49	85.47	64.15	85.40
MobileNetV2_0.25	50.74	74.56	50.14	74.13

vgg

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
VGG11	66.62	87.34	67.26	87.73
VGG13	67.74	88.11	68.15	88.47
VGG16	73.23	91.31	70.09	89.70
VGG19	74.11	91.35	70.86	90.17
VGG11_bn	68.59	88.72	68.94	88.88
VGG13_bn	68.84	88.82	69.51	89.46
VGG16_bn	73.10	91.76	72.07	90.97
VGG19_bn	74.33	91.85	72.85	91.26

Note: the vgg model here is converted from torchvision

resnext

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
ResNext50_32x4d	79.32	94.53	79.41	94.54
ResNext101_32x4d	80.37	95.06	80.52	95.20
ResNext101_64x4d	80.69	95.17	80.84	95.27
SE_ResNext50_32x4d	79.95	94.93	80.17	94.97
SE_ResNext101_32x4d	80.91	95.39	81.27	95.42
SE_ResNext101_64x4d	81.01	95.32	81.19	95.60

resnetv1b_pruned

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
resnet18_v1b_0.89	67.2	87.45	65.78	86.63
resnet50_v1d_0.86	78.02	93.82	77.61	93.90
resnet50_v1d_0.48	74.66	92.34	74.10	92.10
resnet50_v1d_0.37	70.71	89.74	69.47	89.12
resnet50_v1d_0.11	63.22	84.79	61.12	83.31
resnet101_v1d_0.76	79.46	94.69	79.55	94.81
resnet101_v1d_0.73	78.89	94.48	78.68	94.41

squeezenet

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
SqueezeNet1.0	56.11	79.09	55.67	78.47
SqueezeNet1.1	54.96	78.17	55.27	78.55

densenet

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
DenseNet121	74.97	92.25	74.65	92.15
DenseNet161	77.70	93.80	77.64	93.97
DenseNet169	76.17	93.17	76.26	93.18
DenseNet201	77.32	93.62	77.64	93.97

inception

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
InceptionV3	78.77	94.39	78.62	94.42

InceptionV3 is evaluated with input size of 299x299.

alexnet

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
AlexNet	54.92	78.03	54.28	77.68

darknet

Model	Acc@1(gluon-cv)	Acc@5(gluon-cv)	Acc@1	Acc@5
darknet53	78.56	94.43	78.54	94.54

TODO

• Add more pretrained models