Table 1.
The details of the proposed initial CNN model
Type | Patch size/Stride | Output | Depth | Params |
---|---|---|---|---|
Convolution | 7 ×7/2 | 112 ×112×64 | 1 | 2.7K |
Max pool | 3 ×3/2 | 56 ×56×64 | 0 | |
Convolution | 1 ×1/1 | 56 ×56×64 | 1 | 0.8K |
Convolution | 3 ×3/1 | 56 ×56×192 | 1 | 112K |
Max pool | 3 ×3/2 | 28 ×28×192 | 0 | |
Inception(3a) | 28 ×28×256 | 2 | 159K | |
SEP block | 1 ×1 | 28 ×28×256 | 2 | 32K |
Inception(3b) | 28 ×28×480 | 2 | 380K | |
SEP block | 1 ×1 | 28 ×28×480 | 2 | 32K |
Max pool | 3 ×3/2 | 14 ×14×480 | 0 | |
Inception(4a) | 14 ×14×512 | 2 | 364K | |
SEP block | 1 ×1 | 14 ×14×512 | 2 | 32K |
Inception(4b) | 14 ×14×512 | 2 | 437K | |
SEP block | 1 ×1 | 14 ×14×512 | 2 | 32K |
Inception(4c) | 14 ×14×512 | 2 | 840K | |
SEP block | 1 ×1 | 14 ×14×512 | 2 | 32K |
Inception(4d) | 14 ×14×528 | 2 | 580K | |
SEP block | 1 ×1 | 14 ×14×528 | 2 | 32K |
Inception(4e) | 14 ×14×1856 | 2 | 840K | |
SEP block | 1 ×1 | 14 ×14×1856 | 2 | 32K |
Max pool | 3 ×3/2 | 7 ×7×1856 | 0 | |
Ave pool | 7 ×7/1 | 1 ×1×1856 | 0 | |
Linear | 1 ×1×2 | 1 | 2K | |
Softmax | 1 ×1×2 | 0 |
The output of the convolution layer and SEP block may change after the channel pruning stage in every model compression loop