Table 6 Layer distribution of execution time and operations for classifier inference
From: A depthwise separable convolutional neural network for keyword spotting on an embedded system
Layer | Execution time [ms] | Millions of operations |
|---|---|---|
Conv 1 | 99.0 (43.5%) | 3.04 (23.2%) |
DW-conv 1 | 8.3 (3.7%) | 0.17 (1.4%) |
PW-conv 1 | 13.0 (5.7%) | 1.50 (11.5%) |
DW-conv 2 | 8.3 (3.7%) | 0.17 (1.4%) |
PW-conv 2 | 13.0 (5.7%) | 1.50 (11.5%) |
DW-conv 3 | 8.3 (3.7%) | 0.17 (1.4%) |
PW-conv 3 | 13.0 (5.7%) | 1.50 (11.5%) |
DW-conv 4 | 8.3 (3.7%) | 0.17 (1.4%) |
PW-conv 4 | 13.0 (5.7%) | 1.50 (11.5%) |
DW-conv 5 | 8.3 (3.7%) | 0.17 (1.4%) |
PW-conv 5 | 13.0 (5.7%) | 1.50 (11.5%) |
DW-conv 6 | 8.3 (3.7%) | 0.17 (1.4%) |
PW-conv 6 | 13.0 (5.7%) | 1.50 (11.5%) |
Avg-pool | 0.5 (0.2%) | 0.01 (0.1%) |
FC | 0.1 (0.04%) | 0.001 (<0.1%) |
Total | 227.4 | 13.12 |