##############################
+class NoiseInjector(nn.Module):
+ def __init__(self):
+ super().__init__()
+ self.noise_std = 0.0
+
+ def forward(self, x):
+ if self.noise_std > 0:
+ x = x + torch.randn(x.size(), device=x.device) * self.noise_std
+ return x
+
+
+def set_noise_injection(model, noise_std):
+ for m in model.modules():
+ if isinstance(m, NoiseInjector):
+ m.noise_std = noise_std
+
+
+##############################
+
+
class MyGPT(nn.Module):
def __init__(
self,
for b in range(nb_blocks):
trunk_blocks += [
WithResidual(
- CacheWrapper(nn.LayerNorm((dim_model,))),
+ CacheWrapper(
+ nn.LayerNorm((dim_model,)),
+ NoiseInjector(),
+ ),
QKVAttention(
dim_in=dim_model,
dim_qk=dim_keys,
WithResidual(
CacheWrapper(
nn.LayerNorm((dim_model,)),
+ NoiseInjector(),
nn.Linear(in_features=dim_model, out_features=dim_hidden),
nn.ReLU(),
nn.Linear(in_features=dim_hidden, out_features=dim_model),