Revert scaling, scale only grad.

egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py

@@ -900,6 +900,26 @@ def with_loss(x, y):
     return WithLoss.apply(x, y)
 
 
+class ScaleGradFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, alpha: float) -> Tensor:
+        ctx.alpha = alpha
+        return x
+    @staticmethod
+    def backward(ctx, grad: Tensor):
+        return grad * ctx.alpha, None
+
+def scale_grad(x: Tensor, alpha: float):
+    return ScaleGradFunction.apply(x, alpha)
+
+class ScaleGrad(nn.Module):
+    def __init__(self, alpha: float):
+        super().__init__()
+        self.alpha = alpha
+
+    def forward(self, x: Tensor) -> Tensor:
+        return scale_grad(x, self.alpha)
+
 class LimitParamValue(torch.autograd.Function):
     @staticmethod
     def forward(ctx, x: Tensor, min: float, max: float):

egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py

@@ -43,6 +43,7 @@ from scaling import (
     ScheduledFloat,
     FloatLike,
     limit_param_value,
+    ScaleGrad,
 )
 from torch import Tensor, nn
 
@@ -1719,25 +1720,22 @@ class Conv2dSubsampling(nn.Module):
         # a too-large gradient).
 
         self.conv = nn.Sequential(
-            ScalarMultiply(0.1),
-            ScaledConv2d(
+            nn.Conv2d(
                 in_channels=1,
                 out_channels=layer1_channels,
                 kernel_size=3,
                 padding=(0, 1),  # (time, freq)
-                initial_scale=5.0,
             ),
-            ScalarMultiply(0.25),
+            ScaleGrad(0.1),
             ActivationBalancer(layer1_channels,
                                channel_dim=1),
             DoubleSwish(),
-            ScaledConv2d(
+            nn.Conv2d(
                 in_channels=layer1_channels,
                 out_channels=layer2_channels,
                 kernel_size=3,
                 stride=2,
                 padding=0,
-                initial_scale=5.0,
             ),
             ActivationBalancer(layer2_channels,
                                channel_dim=1),