Table 3.

Architecture for proposed fine-tuned ResNet50 TL. Building blocks are shown in brackets, with the numbers of blocks stacked. Downsampling is performed by conv3_1, conv4_1, and conv5_1 with a stride of 2.

Layer name	Output size	Layer
conv1	112 × 112	7 × 7, 64, stride 2
conv2_x	56 × 56	3 × 3 max pool, stride 2
		[1 × 1, 64 3 × 3, 64 1 × 1, 256] $\times 3$
conv3_x	28 × 28	[1 × 1, 128 3 × 3, 128 1 × 1, 512] $\times 4$
conv4_x	$14\times 14$	[$1\times 1$, 256 3 × 3, 256 1 × 1, 1024] $\times 6$
conv5_x	7 × 7	[1 × 1, 512 3 × 3, 512 1 × 1, 2048] $\times 3$
fc1	1 × 1	Average pool
		in_features $=$ 2048, out_features $=$ 2048
fc2	1 × 1	Dropout 0.5
		in_features $=$ 2048, out_features $=$ 2048
fc3	1 × 1	relu, dropout 0.5
		in_features $=$ 2048, out_features=2