fix bugs

fairseq/models/speech_to_text/s2t_transformer.py

@@ -618,6 +618,11 @@ class S2TTransformerModel(FairseqEncoderDecoderModel):
             help="share the weight of target ctc and embed",
         )
         parser.add_argument(
+            "--share-xctc-and-ctc",
+            action="store_true",
+            help="share the weight of target ctc and ctc",
+        )        
+        parser.add_argument(
             "--inter-xctc-layers",
             default=None,
             type=str,
@@ -1197,7 +1202,7 @@ class S2TTransformerEncoder(FairseqEncoder):
                     xctc_norm = LayerNorm(dim)
                     setattr(self, "xctc_norm%d" % layer_idx, xctc_norm)
 
-                # consider layer norm
+            # consider layer norm
             if not hasattr(self, "xctc"):
                 self.xctc = CTC(
                     dim,
@@ -1212,6 +1217,8 @@ class S2TTransformerEncoder(FairseqEncoder):
                     == embed_tokens.weight.size()
                 ):
                     self.xctc.ctc_projection.weight = embed_tokens.weight
+                elif getattr(self, "share_xctc_and_ctc", False) and hasattr(self, "ctc"):
+                    self.xctc.ctc_projection.weight = self.ctc.ctc_projection.weight
 
             strategy = {
                 "embed_norm": getattr(args, "pae_embed_norm", False),
@@ -2365,6 +2372,7 @@ def base_architecture(args):
     # XCTC
     args.xctc_layer = getattr(args, "xctc_layer", 0)
     args.share_xctc_and_embed = getattr(args, "share_xctc_and_embed", False)
+    args.share_xctc_and_ctc = getattr(args, "share_xctc_and_ctc", False)
     args.xctc_pae = getattr(args, "xctc_pae", args.ctc_pae)
     args.axctc_pae = getattr(args, "axctc_pae", args.xctc_pae)
     args.share_pae_and_xctc = getattr(args, "share_pae_and_xctc", False)