GAT: Attention Passo a Passo
1
2
3
4
5
Passo 4 de 5
1. Embeddings ← Slide 16
hv=[1,1] hA=[2,2]
hB=[1,0] hC=[0,2]
← Ver exemplo numérico completo no Slide 16
hB=[1,0] hC=[0,2]
2. Projeção W ← Slide 16
W=[[1,0],[0,2]]
Whv=[1,2] WhA=[2,4]
WhB=[1,0] WhC=[0,4]
← Ver cálculo matricial no Slide 16
Whv=[1,2] WhA=[2,4]
WhB=[1,0] WhC=[0,4]
3. Score com vetor a ← Slide 16
a=[1,1,1,0]
e(v,A)=5 e(v,B)=4 e(v,C)=3
← Ver tabela de scores no Slide 16
e(v,A)=5 e(v,B)=4 e(v,C)=3
Passos 1–3 calculados no Slide 16. Aqui usamos os α resultantes.
4. Softmax → α
αvu = exp(evu) / Σk exp(evk)
5. Agregação Ponderada
h'v = σ( αA·WhA + αB·WhB + αC·WhC )
0.67 ×
| 2 | 4 |
+
0.24 ×
| 1 | 0 |
+
0.09 ×
| 0 | 4 |
=
σ(
) ≈
| 1.58 | 3.04 |
| 0.83 | 0.95 |
🔑 W aqui é a mesma do Slide 16: W=[[1,0],[0,2]]