Table 5.

Imagenet network topologies for conversion training.

Model	Configuration	BackRes
VGG16	Input–Conv1(3,64,3x3/1)–Conv2(64,64,3x3/1)–	Not applicable
	–Pool(2 × 2/2)–Conv3(64,128,3 × 3/1)–Pool(2 × 2/2)–Conv4(128,256,3 × 3/1)–
	–Conv5(256,256,3 × 3/1)–Conv6(256,256,3 × 3/1)–Pool(2 × 2/2)–Conv7(256,512,3 × 3/1)–
	–Conv8(512,512,3 × 3/1)–Conv9(512,512,3 × 3/1)–Conv10(512,512,3 × 3/1)–Conv11(512,512,3 × 3/1)–
	–Conv12(512,512,3 × 3/1)–Conv13(512,512,3 × 3/1)–Pool(2 × 2/2)–Pool(2 × 2/2)–
	–FC1(25088,4096)–FC2(4096,1000)
VGG11x2	Input–Conv1(3,64,3 × 3/1)–Conv2(64,64,3 × 3/1)–	$\text{[Conv5]}$& $\text{[Conv7-Conv8-Conv9]}$ repeated 2 times
	–Pool(2 × 2/2)–Conv3(64,128,3 × 3/1)–Pool(2 × 2/2)–Conv4(256,256,3 × 3/1)–$\text{Conv5}(\mathrm{256,256,3}\times \mathrm{3/1})$–
	–$\text{Conv5}(\mathrm{256,256,3}\times \mathrm{3/1})$–Pool(2 × 2/2)–Conv6(256,512,3 × 3/1)–$\text{Conv7}(\mathrm{512,512,3}\times \mathrm{3/1})$–
	–$\text{Conv8}(\mathrm{512,512,3}\times \mathrm{3/1})$$\text{–Conv9}(\mathrm{512,512,3}\times \mathrm{3/1})$–$\text{Conv7}(\mathrm{512,512,3}\times \mathrm{3/1})$–$\text{Conv8}(\mathrm{512,512,3}\times \mathrm{3/1})$–
	–$\text{Conv9}(\mathrm{512,512,3}\times \mathrm{3/1})$–Pool(2 × 2/2)–Pool(2 × 2/2)–
	–FC1(25088,4096)–FC2(4096,1000)

Notations are same as that of Table 3. Layers with BackRes connections and repeated computations have been highlighted in red.