Table A1.

Detailed structure of models except for mobile models. SA stands for Soft-Attention, SA Module denotes whether that model uses the Soft-Attention module. GAP stands for Global Average Pooling. FC stands for Fully Connected Layer.

DenseNet-201	DenseNet-201 + SA	Inception-ResNetV2	Inception-ResNetV2 + SA	ResNet-50	ResNet-50 + SA	ResNet-152	ResNet-152 + SA	NasNet-Large	NasNet-Large + SA
Conv2D $7\times 7$	Conv2D $7\times 7$	STEM	STEM	Conv2D $7\times 7$	Conv2D $7\times 7$	Conv2D $7\times 7$	Conv2D $7\times 7$	Conv2D $3\times 3$	Conv2D $3\times 3$
Pooling $3\times 3$	Pooling $3\times 3$			Pooling $3\times 3$	Pooling $3\times 3$	Pooling $3\times 3$	Pooling $3\times 3$	Pooling	Pooling
DenseBlock × 6	DenseBlock × 6	Inception ResNet A × 10	Inception ResNet A × 10	Residual Block × 3	Residual Block × 3	Residual Block × 3	Residual Block × 3	Reduction Cell × 2	Reduction Cell × 2
Conv2D $1\times 1$	Conv2D $1\times 1$	Reduction A	Reduction A					Normal Cell × N	Normal Cell × N
Average pool $2\times 2$	Average pool $2\times 2$
DenseBlock × 12	DenseBlock × 12	Inception ResNet B × 20	Inception ResNet B × 20	Residual Block × 4	Residual Block × 4	Residual Block × 8	Residual Block × 8	Reduction Cell	Reduction Cell
Conv2D $1\times 1$	Conv2D $1\times 1$	Reduction B	Reduction B					Normal Cell × N	Normal Cell × N
Average pool $2\times 2$	Average pool $2\times 2$
DenseBlock × 48	DenseBlock × 12	Inception ResNet C × 5	Inception ResNet C × 5	Residual Block × 6	Residual Block × 6	Residual Block × 36	Residual Block × 36	Reduction Cell	Reduction Cell
Conv2D $1\times 1$	Conv2D $1\times 1$							Normal Cell × N	Normal Cell × N-2
Average pool $2\times 2$	Average pool $2\times 2$
DenseBlock × 29	DenseBlock × 29			Residual Block × 3		Residual Block × 3
DenseBlock × 3	SA Module		SA Module		SA Module		SA Module		SA Module
GAP $7\times 7$		Average pool		GAP $7\times 7$		GAP $7\times 7$
FC 1000D		Dropout (0.8)		FC 1000D		FC 1000D
SoftMax	SoftMax	SoftMax	SoftMax	SoftMax	SoftMax	SoftMax	SoftMax	SoftMax	SoftMax