Nongkai Tian
Nongkai Tian
对EfficientNetB0到EfficientNetB7参数的详细定义在[EfficientNet网络详解](https://blog.csdn.net/qq_37541097/article/details/114434046)中。根据[EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/pdf/1905.11946.pdf)中的公式3,可以得知在EfficientNetB7中, _r=γ^Φ=1.15^Φ=600/224=2.6786_ _Φ=log(2.6786)/log(1.15)=7.04_ 进而可以推出 _depth_coefficient=d=α^Φ=1.2^7.04=3.6094≠3.1_ 我阅读了若干个关于EfficientNet的仓库的实现代码,发现几乎所有的实现代码都对EfficientNetB7中的depth_coefficient设置为3.1,所以我不确定上述推导过程是否正确。如果错误,正确的推导过程是什么呢?
The code is found on line 475 of [EfficientNet-PyTorch/efficientnet_pytorch/utils.py](https://github.com/lukemelas/EfficientNet-PyTorch/blob/master/efficientnet_pytorch/utils.py). According to formula 3 of the original paper, the image size of efficientnet-b7 is 600. _r=γ^Φ=1.15^Φ=600/224=2.6786_, so _Φ=log(2.6786)/log(1.15)=7.04_. Therefore, _depth=d=α^Φ=1.2^7.04=3.6094≠3.1_ I...