Wireless Network Optimization for Federated Learning with Model Compression in Hybrid VLC/RF Systems

. 2021 Oct 27;23(11):1413. doi: 10.3390/e23111413

Algorithm 1: FL Model Compression

1:
Input: $W^{I}$
2:
for $τ \in {1, \dots, T}$ do
3:
for $n \in {1, \dots, N}$ do
4:
Client n does:
5:
$(W_{C}^{t - 1}, M) \leftarrow D o w n l o a d_{B S \to n} (W_{C}^{t - 1})$
6:
$W_{n}^{t} \leftarrow W_{C}^{t - 1} + (1 - M) \otimes W_{n}^{t - 1}$
7:
$G_{n}^{t} \leftarrow T r a i n (W_{n}^{t}, D_{n}) + R_{n}^{t - 1} - W_{n}^{t}$
8:
$M_{n} \leftarrow C o m p r e s s (G_{n}^{t})$
9:
$G_{n, C}^{t} \leftarrow G_{n}^{t} \otimes M_{n}$
10:
$W_{n, C}^{t} \leftarrow W_{n}^{t} \otimes M_{n}$
11:
$R_{n}^{t} \leftarrow G_{n}^{t} - G_{n, C}^{t}$
12:
$S a v e (R_{n}^{t}, W_{n}^{t})$
13:
$U p l o a d_{n \to B S} (W_{n, C}^{t})$
14:
end for
15:
BS does:
16:
for $n \in S = S_{1} \cup S_{2}$ do
17:
$(W_{n, C}^{t}, M_{n}) \leftarrow W_{n, C}^{t}$
18:
end for
19:
$W_{C}^{t} \leftarrow A g g r e g a t e (\frac{1}{| S |} \sum_{n \in S} W_{n, C}^{t}) + R^{t - 1}$
20:
$W^{t} \leftarrow W_{C}^{t} + (1 - M_{1}) \otimes (1 - M_{2}) \dots \otimes (1 - M_{n}) \otimes W^{t - 1}$
21:
$G^{t} \leftarrow W^{t} - W^{t - 1}$
22:
$M \leftarrow C o m p r e s s (G^{t})$
23:
$G_{C}^{t} \leftarrow G^{t} \otimes M$
24:
$W_{C}^{t} \leftarrow W^{t} \otimes M$
25:
$R^{t} \leftarrow G^{t} - G_{C}^{t}$
26:
$S a v e (R^{t}, W^{t})$
27:
$T r a n s m i t s_{B S \to n} (W_{C}^{t})$
28:
end for
29:
Return $W^{F}$