Table 5 Single-inference accuracies of networks with individually quantized network parts
From: A depthwise separable convolutional neural network for keyword spotting on an embedded system
Floating point accuracy 83.2 % (79.3 %) | |||
|---|---|---|---|
Bit width | 8-bit | 4-bit | 2-bit |
Convolution weights | 83.2 % (79.1 %) | 82.3 % (78.8 %) | 46.8 % (44.6 %) |
DW-Convolution weigths | 83.3 % (79.3 %) | 80.0 % (76.7 %) | 12.0 % (12.1 %) |
PW-Convolution weights | 83.4 % (79.2 %) | 78.9 % (75.4 %) | 11.2 % (10.8 %) |
Fully connected weights | 83.2 % (79.3 %) | 82.9 % (79.1 %) | 70.1 % (63.9 %) |
Layer activations | 83.2 % (79.2 %) | 53.6 % (52.0 %) | 8.6 % (8.7 %) |