keras efficientnet introduction


Guide

About EfficientNet Models

params
flops

compared with resnet50, EfficientNet-B4 improves the top-1 accuracy from 76.3% of ResNet-50 to 82.6% (+6.3%), under similar FLOPS constraint.

Using Pretrained EfficientNet Checkpoints

b0-b7 top-1 on imagenet

Keras Models Performance

  • The top-k errors were obtained using Keras Applications with the TensorFlow backend on the 2012 ILSVRC ImageNet validation set and may slightly differ from the original ones.

The input size used was 224x224 for all models except NASNetLarge (331x331), InceptionV3 (299x299), InceptionResNetV2 (299x299), Xception (299x299),
EfficientNet-B0 (224x224), EfficientNet-B1 (240x240), EfficientNet-B2 (260x260), EfficientNet-B3 (300x300), EfficientNet-B4 (380x380), EfficientNet-B5 (456x456), EfficientNet-B6 (528x528), and EfficientNet-B7 (600x600).

notice

  • Top-1: single center crop, top-1 error
  • Top-5: single center crop, top-5 error
  • 10-5: ten crops (1 center + 4 corners and those mirrored ones), top-5 error
  • Size: rounded the number of parameters when include_top=True
  • Stem: rounded the number of parameters when include_top=False
Top-1Top-510-5SizeStemReferences
VGG1628.7329.9508.834138.4M14.7M[paper] [tf-models]
VGG1928.74410.0128.774143.7M20.0M[paper] [tf-models]
ResNet5025.0727.9406.82825.6M23.6M[paper] [tf-models] [torch] [caffe]
ResNet10123.5807.2146.09244.7M42.7M[paper] [tf-models] [torch] [caffe]
ResNet15223.3966.8825.90860.4M58.4M[paper] [tf-models] [torch] [caffe]
ResNet50V224.0406.9665.89625.6M23.6M[paper] [tf-models] [torch]
ResNet101V222.7666.1845.15844.7M42.6M[paper] [tf-models] [torch]
ResNet152V221.9685.8384.90060.4M58.3M[paper] [tf-models] [torch]
ResNeXt5022.2606.1905.41025.1M23.0M[paper] [torch]
ResNeXt10121.2705.7064.84244.3M42.3M[paper] [torch]
InceptionV322.1026.2805.03823.9M21.8M[paper] [tf-models]
InceptionResNetV219.7444.7483.96255.9M54.3M[paper] [tf-models]
Xception20.9945.5484.73822.9M20.9M[paper]
MobileNet(alpha=0.25)48.41824.20821.1960.5M0.2M[paper] [tf-models]
MobileNet(alpha=0.50)35.70814.37612.1801.3M0.8M[paper] [tf-models]
MobileNet(alpha=0.75)31.58811.7589.8782.6M1.8M[paper] [tf-models]
MobileNet(alpha=1.0)29.57610.4968.7744.3M3.2M[paper] [tf-models]
MobileNetV2(alpha=0.35)39.91417.56815.4221.7M0.4M[paper] [tf-models]
MobileNetV2(alpha=0.50)34.80613.93811.9762.0M0.7M[paper] [tf-models]
MobileNetV2(alpha=0.75)30.46810.8249.1882.7M1.4M[paper] [tf-models]
MobileNetV2(alpha=1.0)28.6649.8588.3223.5M2.3M[paper] [tf-models]
MobileNetV2(alpha=1.3)25.3207.8786.7285.4M3.8M[paper] [tf-models]
MobileNetV2(alpha=1.4)24.7707.5786.5186.2M4.4M[paper] [tf-models]
DenseNet12125.0287.7426.5228.1M7.0M[paper] [torch]
DenseNet16923.8246.8245.86014.3M12.6M[paper] [torch]
DenseNet20122.6806.3805.46620.2M18.3M[paper] [torch]
NASNetLarge17.5023.9963.41293.5M84.9M[paper] [tf-models]
NASNetMobile25.6348.1466.7587.7M4.3M[paper] [tf-models]
EfficientNet-B022.8106.5085.8585.3M4.0M[paper] [tf-tpu]
EfficientNet-B120.8665.5525.0507.9M6.6M[paper] [tf-tpu]
EfficientNet-B219.8205.0544.5389.2M7.8M[paper] [tf-tpu]
EfficientNet-B318.4224.3243.90212.3M10.8M[paper] [tf-tpu]
EfficientNet-B417.0403.7403.34419.5M17.7M[paper] [tf-tpu]
EfficientNet-B516.2983.2903.11430.6M28.5M[paper] [tf-tpu]
EfficientNet-B615.9183.1022.91643.3M41.0M[paper] [tf-tpu]
EfficientNet-B715.5703.1602.90666.7M64.1M[paper] [tf-tpu]

Reference

History

  • 20190912: created.

Author: kezunlin
Reprint policy: All articles in this blog are used except for special statements CC BY 4.0 reprint polocy. If reproduced, please indicate source kezunlin !
评论
  TOC