Table 9 Ablation on architecture hyperparameters using UTD-MHAD.

From: A tiny inertial transformer for human activity recognition via multimodal knowledge distillation and explainable AI

Configuration

Accuracy (%)

Model size (MB)

FLOPs (M)

\(L = 1\) (D=128, P=20)

96.13

1.43

6.4

\(L = 2\) (default)

98.71

2.45

11.3

\(L = 3\)

98.40

3.36

16.5

\(D = 64\) (L=2, P=20)

96.89

1.61

7.8

\(D = 128\) (default)

98.71

2.45

11.3

\(D = 256\)

98.83

4.28

19.1

\(P = 10\) (L=2, D=128)

96.97

2.45

11.2

\(P = 20\) (default)

98.71

2.45

11.3

\(P = 40\)

98.55

2.45

11.9

Back to article page