grad_compute(): pre-multiply weight decay with learning rate

2026-01-05 19:40:05 +01:00
parent 10999ec731
commit d238d2954e
1 changed files with 1 additions and 1 deletions
@@ -98,7 +98,7 @@ class Entity:
 		# Compute gradient
 		self.grad.b_v = (momentum * self.grad.b_v + learning_rate * d_bv)
 		self.grad.b_h = (momentum * self.grad.b_h + learning_rate * d_bh)
-		self.grad.w_hv = (momentum * self.grad.w_hv + learning_rate * d_whv - weight_decay * self.state.w_hv)
+		self.grad.w_hv = (momentum * self.grad.w_hv + learning_rate * d_whv - learning_rate * weight_decay * self.state.w_hv)

 		return self.grad