1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
|
// Forward Deep Neural Net (DNN), any number of layers of any size each
declare name "DNN";
declare author "JOS";
declare license "STK-4.3";
import("stdfaust.lib");
layerSizes = (8,5,8); // autoencoder with 8 in & out, 5-state hidden layer
w(m,n,k) = m*100+n*10+k; // placeholder weights: m=layer, n=fromNode, k=destNode
M = ba.count(layerSizes);
N(l) = ba.take(l+1,layerSizes); // Nodes per layer
process = seq(m, M-1, layer(m))
// look at weights:
// process = par(m,M,par(n,N(m),par(k,N(m),w(m,n,k))))
with {
layer(m) = weights(m) :> nonlinearities(m);
nonlinearities(m) = bus(N(m)*N(m+1)) :> par(n,N(m+1),nl(n));
weights(m) = bus(N(m)) <: par(n,N(m),(bus(N(m+1))<:wts(m,n)));
wts(m,n) = bus(N(m+1)) : par(k,N(m+1),*(w(m,n,k)));
nl(n,x) = x * (x>0); // ReLU
bus(N) = par(k,N,_);
};
|
