Complete Architectural Details of all EfficientNet Models

Original article was published on Deep Learning on Medium

It’s easy to see the difference among all the models and they gradually increased the number of sub-blocks. If you understood the architectures I will encourage you to take any model and print its summary and have a go through it to know it more thoroughly. The table shown below denotes the kernel size for convolution operations along with the resolution, channels, and layers in EfficientNet-B0.

Kernel Size, resolution, channels, and no. of layers information.

This table was included in the original paper. The resolution remains the same as for the whole family. I don’t know about whether the kernel size changes or remains the same so if anyone knows leave a reply. The number of layers is already shown above in the figures. The number of channels varies and it is calculated from the information seen from each model’s summary and is presented below.

╔═══════╦══════╦══════╦══════╦══════╦══════╦══════╦══════╗
║ Stage ║ B1 ║ B2 ║ B3 ║ B4 ║ B5 ║ B6 ║ B7 ║
╠═══════╬══════╬══════╬══════╬══════╬══════╬══════╬══════╣
║ 1 ║ 32 ║ 32 ║ 40 ║ 48 ║ 48 ║ 56 ║ 64 ║
║ 2 ║ 16 ║ 16 ║ 24 ║ 24 ║ 24 ║ 32 ║ 32 ║
║ 3 ║ 24 ║ 24 ║ 32 ║ 32 ║ 40 ║ 40 ║ 48 ║
║ 4 ║ 40 ║ 48 ║ 48 ║ 56 ║ 64 ║ 72 ║ 80 ║
║ 5 ║ 80 ║ 88 ║ 96 ║ 112 ║ 128 ║ 144 ║ 160 ║
║ 6 ║ 112 ║ 120 ║ 136 ║ 160 ║ 176 ║ 200 ║ 224 ║
║ 7 ║ 192 ║ 208 ║ 232 ║ 272 ║ 304 ║ 344 ║ 384 ║
║ 8 ║ 320 ║ 352 ║ 384 ║ 448 ║ 512 ║ 576 ║ 640 ║
║ 9 ║ 1280 ║ 1408 ║ 1536 ║ 1792 ║ 2048 ║ 2304 ║ 2560 ║
╚═══════╩══════╩══════╩══════╩══════╩══════╩══════╩══════╝

Medium does not has any format to make tables, so if you want to create tables like the one above you create ASCII tables from this site.