Introduce alpha for DoubleSwish, set it to -0.05.

egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py

@@ -1064,7 +1064,8 @@ class MaxEig(torch.nn.Module):
 
 class DoubleSwishFunction(torch.autograd.Function):
     """
-      double_swish(x) = x * torch.sigmoid(x-1)
+      double_swish(x) = x * (torch.sigmoid(x-1) + alpha)
+    for e.g. alpha=-0.05 (user supplied).
     This is a definition, originally motivated by its close numerical
     similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
 
@@ -1079,9 +1080,10 @@ class DoubleSwishFunction(torch.autograd.Function):
     """
 
     @staticmethod
-    def forward(ctx, x: Tensor) -> Tensor:
+    def forward(ctx, x: Tensor, alpha: float) -> Tensor:
         requires_grad = x.requires_grad
         x_dtype = x.dtype
+        ctx.alpha = alpha
         if x.dtype == torch.float16:
             x = x.to(torch.float32)
 
@@ -1105,6 +1107,7 @@ class DoubleSwishFunction(torch.autograd.Function):
                 assert d_scaled.max() < 256.0
             d_int = d_scaled.to(torch.uint8)
             ctx.save_for_backward(d_int)
+        y = y + alpha * x
         if x.dtype == torch.float16 or torch.is_autocast_enabled():
             y = y.to(torch.float16)
         return y
@@ -1112,20 +1115,29 @@ class DoubleSwishFunction(torch.autograd.Function):
     @staticmethod
     def backward(ctx, y_grad: Tensor) -> Tensor:
         d, = ctx.saved_tensors
+        alpha = ctx.alpha
         # the same constants as used in forward pass.
         floor = -0.043637
         ceil = 1.2
         d = (d * ((ceil - floor) / 255.0) + floor)
-        return (y_grad * d)
+        return (y_grad * (d + alpha)), None
 
 class DoubleSwish(torch.nn.Module):
+    def __init__(self,
+                 alpha: float = -0.05):
+        super().__init__()
+        self.alpha = alpha
+
+    def extra_repr(self) -> str:
+        return 'alpha={}'.format(self.alpha)
+
     def forward(self, x: Tensor) -> Tensor:
         """Return double-swish activation function which is an approximation to Swish(Swish(x)),
         that we approximate closely with x * sigmoid(x-1).
         """
         if torch.jit.is_scripting():
-            return x * torch.sigmoid(x - 1.0)
-        return DoubleSwishFunction.apply(x)
+            return x * (torch.sigmoid(x - 1.0) + self.alpha)
+        return DoubleSwishFunction.apply(x, self.alpha)
 
 
 class TanSwishFunction(torch.autograd.Function):