Skip to content

Latest commit

 

History

History
139 lines (107 loc) · 11.7 KB

README.md

File metadata and controls

139 lines (107 loc) · 11.7 KB

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