Table 3. VGG configuration for ImageNet100 (Simonyan & Zisserman, 2015).

Layer	Output size	Network configuration
conv-1	$224\times 224$	$\left[3\times 3,64\right]\times 2$
	$112\times 112$	maxpool
conv-2	$112\times 112$	$\left[3\times 3,64\right]\times 2$
	$56\times 56$	maxpool
conv-3	$56\times 56$	$\left[3\times 3,128\right]\times 1$
	$28\times 28$	maxpool
conv-4	$28\times 28$	$\left[3\times 3,128\right]\times 1$
	$14\times 14$	maxpool
conv-5	$14\times 14$	$\left[3\times 3,128\right]\times 1$
	$7\times 7$	maxpool
conv-6	$7\times 7$	$\left[3\times 3,128\right]\times 1$
	2 × 2	AdaptiveAvgPool
	1 × 1	FC-100

Notes:

$[m\times m,n]\times k$, $m\times m$: convolution kernel size, n: output channels; k: the repeat number of the layer.

Each convolutional layer is followed by a BN and ReLU layer.

maxpool: $2\times 2$ with stride 2.

Output size: $width\times height$.