Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b0_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b1_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b2_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b3_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b4_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b5_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b6_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b7_ap |
SHOW MORE |
Training Techniques | AdvProp, RMSProp, Weight Decay, Label Smoothing, AutoAugment, Stochastic Depth |
---|---|
Architecture | 1x1 Convolution, Average Pooling, Convolution, Dense Connections, Dropout, Inverted Residual Block, Batch Normalization, Squeeze-and-Excitation Block, Swish |
ID | tf_efficientnet_b8_ap |
SHOW MORE |
AdvProp is an adversarial training scheme which treats adversarial examples as additional examples, to prevent overfitting. Key to the method is the usage of a separate auxiliary batch norm for adversarial examples, as they have different underlying distributions to normal examples.
To load a pretrained model:
import timm
m = timm.create_model('tf_efficientnet_b0_ap', pretrained=True)
m.eval()
Replace the model name with the variant you want to use, e.g. tf_efficientnet_b0_ap
. You can find the IDs in the model summaries at the top of this page.
You can follow the timm recipe scripts for training a new model afresh.
@misc{xie2020adversarial,
title={Adversarial Examples Improve Image Recognition},
author={Cihang Xie and Mingxing Tan and Boqing Gong and Jiang Wang and Alan Yuille and Quoc V. Le},
year={2020},
eprint={1911.09665},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
MODEL | TOP 1 ACCURACY | TOP 5 ACCURACY |
---|---|---|
tf_efficientnet_b8_ap | 85.37% | 97.3% |
tf_efficientnet_b7_ap | 85.12% | 97.25% |
tf_efficientnet_b6_ap | 84.79% | 97.14% |
tf_efficientnet_b5_ap | 84.25% | 96.97% |
tf_efficientnet_b4_ap | 83.26% | 96.39% |
tf_efficientnet_b3_ap | 81.82% | 95.62% |
tf_efficientnet_b2_ap | 80.3% | 95.03% |
tf_efficientnet_b1_ap | 79.28% | 94.3% |
tf_efficientnet_b0_ap | 77.1% | 93.26% |