Table 1.

The architecture of the Inception V3 model. Each row represents a layer of the network, and the input of a particular layer is the output of the previous layer. For the inception layers, each row represents parallel sub-layers that were concatenated prior to being passed to the subsequent layer. The list of parameters in each row represents serial processes within the sub-layer. With the Inception 6 layer, the parameters within the parentheses represent additional parallel sub-layers within the serial processes. The number of times each Inception layer is repeated is prepended to each layer name

Type	Patch size/strides	Input size
Conv 1	3 × 3/2	300 × 300 × 1
Conv 2	3 × 3/1	149 × 149 × 32
Conv 3	3 × 3/1	147 × 147 × 32
Max pool 1	3 × 3/2	147 × 147 × 64
Conv 4	1 × 1/1	73 × 73 × 64
Conv 5	3 × 3/1	73 × 73 × 80
Max pool 2	3 × 3/2	71 × 71 × 192
3× Inception 1	1 × 1/1	35 × 35 × 192
	1 × 1/1, 3 × 3/1
	1 × 1/1, 3 × 3/1, 3 × 3/1
	Avg pool 3 × 3/1, 1 × 1/1
Inception 2	3 × 3/2	35 × 35 × 288
	Max pool 3 × 3/2
	1 × 1/1, 3 × 3/1, 3 × 3/2
4× Inception 3	1 × 1/1	17 × 17 × 768
	1 × 1/1, 1 × 7/1, 7 × 1/1
	1 × 1/1, 1 × 7/1, 7 × 1/1, 1 × 7/1, 7 × 1/1
	Avg pool 3 × 3/1, 1 × 1/1
Auxiliary	Avg pool 5 × 5/3, 1 × 1/1, linear, softmax	17 × 17 × 768
Inception 5	1 × 1/1, 3 × 3/2	17 × 17 × 768
	1 × 1/1, 1 × 7/1, 7 × 1/1, 3 × 3/2
	Max pool 3 × 3/2
2× Inception 6	1 × 1/1	8 × 8 × 1280
	1 × 1/1, (1 × 3/1, 3 × 1/1)
	1 × 1/1, 3 × 3/1, (1 × 3/1, 3 × 1/1)
	Avg pool 3 × 3/1, 1 × 1/1
Avg pool	8 × 8/1	8 × 8 × 2048
Output	Dropout, linear, softmax	1 × 1 × 2048