Add baseline for the PeLU expt, keeping only the small normalization-related changes.

egs/librispeech/ASR/conformer_ctc/subsampling.py

@@ -47,11 +47,11 @@ class Conv2dSubsampling(nn.Module):
             nn.Conv2d(
                 in_channels=1, out_channels=odim, kernel_size=3, stride=2
             ),
-            PeLU(cutoff=-1.0),
+            nn.ReLU(),
             nn.Conv2d(
                 in_channels=odim, out_channels=odim, kernel_size=3, stride=2
             ),
-            PeLU(cutoff=-5.0),
+            nn.ReLU(),
         )
         self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
         self.out_norm = nn.LayerNorm(odim, elementwise_affine=False)

egs/librispeech/ASR/transducer_stateless/conformer.py

@@ -156,14 +156,14 @@ class ConformerEncoderLayer(nn.Module):
 
         self.feed_forward = nn.Sequential(
             nn.Linear(d_model, dim_feedforward),
-            PeLU(),
+            Swish(),
             nn.Dropout(dropout),
             nn.Linear(dim_feedforward, d_model),
         )
 
         self.feed_forward_macaron = nn.Sequential(
             nn.Linear(d_model, dim_feedforward),
-            PeLU(),
+            Swish(),
             nn.Dropout(dropout),
             nn.Linear(dim_feedforward, d_model),
         )
@@ -882,7 +882,7 @@ class ConvolutionModule(nn.Module):
             padding=0,
             bias=bias,
         )
-        self.activation = PeLU()
+        self.activation = Swish()
 
     def forward(self, x: Tensor) -> Tensor:
         """Compute convolution module.

egs/librispeech/ASR/transducer_stateless/train.py

@@ -110,7 +110,7 @@ def get_parser():
     parser.add_argument(
         "--exp-dir",
         type=str,
-        default="transducer_stateless/specaugmod_baseline_randcombine1_pelu",
+        default="transducer_stateless/specaugmod_baseline_randcombine1_pelu_base",
         help="""The experiment dir.
         It specifies the directory where all training related
         files, e.g., checkpoints, log, etc, are saved