Skip to content

Latest commit

 

History

History
87 lines (67 loc) · 3.37 KB

README_FixEfficientNet.md

File metadata and controls

87 lines (67 loc) · 3.37 KB
Raw

FixEfficientNet

FixRes is a simple method for fixing the train-test resolution discrepancy. It can improve the performance of any convolutional neural network architecture.

The method is described in "Fixing the train-test resolution discrepancy" (links: arXiv, NeurIPS).

BibTeX reference to cite, if you use it:

@inproceedings{touvron2019FixRes,
       author = {Touvron, Hugo and Vedaldi, Andrea and Douze, Matthijs and J{\'e}gou, Herv{\'e}},
       title = {Fixing the train-test resolution discrepancy},
       booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
       year = {2019},
}
@misc{touvron2020FixEfficientNet,
       author = {Touvron, Hugo and Vedaldi, Andrea and Douze, Matthijs and J{\'e}gou, Herv{\'e}},
       title = {Fixing the train-test resolution discrepancy: FixEfficientNet},
       journal={arXiv preprint arXiv:2003.08237},
       year = {2020},
}

Please notice that our models depend on previous trained models, see References to other models

ImageNet Results

Models Resolution #Parameters Top-1 / Top-5 Extra training data
FixEfficientNet-B0 320 5.3M 79.3 / 94.6
FixEfficientNet-B0 320 5.3M 80.2 / 95.4 x
FixEfficientNet-B1 384 7.8M 81.3 / 95.7
FixEfficientNet-B1 384 7.8M 82.6 / 96.4 x
FixEfficientNet-B2 420 9.2M 82.0 / 96.0
FixEfficientNet-B2 420 9.2M 83.6 / 96.9 x
FixEfficientNet-B3 472 12M 83.0 / 96.4
FixEfficientNet-B3 472 12M 85.0 / 97.4 x
FixEfficientNet-B4 512 19M 84.0 / 97.0
FixEfficientNet-B4 472 19M 85.9 / 97.7 x
FixEfficientNet-B5 576 30M 84.7 / 97.2
FixEfficientNet-B5 576 30M 86.4/ 97.9 x
FixEfficientNet-B6 576 43M 84.9 / 97.3
FixEfficientNet-B6 680 43M 86.7 / 98.0 x
FixEfficientNet-B7 632 66M 85.3 / 97.4
FixEfficientNet-B7 632 66M 87.1 / 98.2 x
FixEfficientNet-B8 800 87.4M 85.7 / 97.6
FixEfficientNet-L2 600 480M 88.5 / 98.7 x

References to other models

Model definition scripts and pretrained weights are from https://github.com/rwightman/pytorch-image-models.

The corresponding papers are:

For models with extra-training data :

@misc{xie2019selftraining,
    author={Qizhe Xie and Minh-Thang Luong and Eduard Hovy and Quoc V. Le,
    title="{Self-training with Noisy Student improves ImageNet classification}",
    journal = {arXiv preprint arXiv:1911.04252},
    year=2019,
}

For models without extra-training data :

@misc{xie2019adversarial,
    author={Cihang Xie and Mingxing Tan and Boqing Gong and Jiang Wang and Alan Yuille and Quoc V. Le,
    title="{Adversarial Examples Improve Image Recognition}",
    journal = {arXiv preprint arXiv:1911.09665},
    year="2019",
}

License

FixRes is CC BY-NC 4.0 licensed, as found in the LICENSE file.