update

toolbox/torchaudio/models/spectrum_dfnet/modeling_spectrum_dfnet.py

@@ -368,6 +368,7 @@ class Encoder(nn.Module):
             kernel_size=config.conv_kernel_size_input,
             bias=False,
             separable=True,
+            fstride=1,
         )
         self.df_conv1 = CausalConv2d(
             in_channels=config.conv_channels,
@@ -386,11 +387,11 @@ class Encoder(nn.Module):
             nn.ReLU(inplace=True)
         )
 
-        # if config.encoder_combine_op == "concat":
-        #     self.embedding_input_size *= 2
-        #     self.combine = Concat()
-        # else:
-        #     self.combine = Add()
+        if config.encoder_combine_op == "concat":
+            self.embedding_input_size *= 2
+            self.combine = Concat()
+        else:
+            self.combine = Add()
 
         # emb_gru
         if config.spec_bins % 8 != 0:
@@ -430,18 +431,18 @@ class Encoder(nn.Module):
         # e2 shape: [batch_size, channels, time_steps, spec_dim // 4]
         # e3 shape: [batch_size, channels, time_steps, spec_dim // 4]
 
-        # # feat_spec, shape: (batch_size, 2, time_steps, df_bins)
-        # c0 = self.df_conv0(feat_spec)
-        # c1 = self.df_conv1(c0)
-        # # c0 shape: [batch_size, channels, time_steps, df_bins]
-        # # c1 shape: [batch_size, channels, time_steps, df_bins // 2]
-        #
-        # cemb = c1.permute(0, 2, 3, 1)
-        # # cemb shape: [batch_size, time_steps, df_bins // 2, channels]
-        # cemb = cemb.flatten(2)
-        # # cemb shape: [batch_size, time_steps, df_bins // 2 * channels]
-        # cemb = self.df_fc_emb(cemb)
-        # # cemb shape: [batch_size, time_steps, spec_dim // 4 * channels]
+        # feat_spec, shape: (batch_size, 2, time_steps, df_bins)
+        c0 = self.df_conv0(feat_spec)
+        c1 = self.df_conv1(c0)
+        # c0 shape: [batch_size, channels, time_steps, df_bins]
+        # c1 shape: [batch_size, channels, time_steps, df_bins // 2]
+
+        cemb = c1.permute(0, 2, 3, 1)
+        # cemb shape: [batch_size, time_steps, df_bins // 2, channels]
+        cemb = cemb.flatten(2)
+        # cemb shape: [batch_size, time_steps, df_bins // 2 * channels]
+        cemb = self.df_fc_emb(cemb)
+        # cemb shape: [batch_size, time_steps, spec_dim // 4 * channels]
 
         # e3 shape: [batch_size, channels, time_steps, spec_dim // 4]
         emb = e3.permute(0, 2, 3, 1)
@@ -449,7 +450,7 @@ class Encoder(nn.Module):
         emb = emb.flatten(2)
         # emb shape: [batch_size, time_steps, spec_dim // 4 * channels]
 
-        # emb = self.combine(emb, cemb)
+        emb = self.combine(emb, cemb)
         # # if concat; emb shape: [batch_size, time_steps, spec_dim // 4 * channels * 2]
         # # if add; emb shape: [batch_size, time_steps, spec_dim // 4 * channels]