control the temperature of the adaptor

fairseq/models/speech_to_text/s2t_sate.py

@@ -60,6 +60,12 @@ class S2TSATEModel(S2TTransformerModel):
             help="adapter type",
         )
         parser.add_argument(
+            "--temperature",
+            default=1.0,
+            type=float,
+            help="temperature of the CTC softmax",
+        )
+        parser.add_argument(
             "--acoustic-encoder",
             default="transformer",
             type=str,
@@ -181,7 +187,7 @@ class Adapter(nn.Module):
             out = torch.mm(distribution.view(-1, embed_dim), self.embed_adapter.weight).view(batch, seq_len, -1)
 
         elif self.adapter_type == "subsample":
-            out = self.subsample_adaptor(x, lengths)
+            out = self.subsample_adaptor(representation, lengths)
 
         elif self.adapter_type == "league":
             linear_out = self.linear_adapter(representation)
@@ -192,6 +198,8 @@ class Adapter(nn.Module):
             soft_out = self.embed_adapter(distribution)
             coef = (self.gate_linear(torch.cat([linear_out, soft_out], dim=-1))).sigmoid()
             out = coef * linear_out + (1 - coef) * soft_out
+        elif self.adapter_type == "none":
+            out = representation
         else:
             out = None
             logging.error("Unsupported adapter type: {}.".format(self.adapter_type))
@@ -256,6 +264,7 @@ class S2TSATEEncoder(FairseqEncoder):
             logging.error("Unsupported model arch {}!".format(acoustic_encoder_type))
 
         # adapter
+        self.temperature = getattr(args, "temperature", 1.0)
         self.adapter = Adapter(args, task.source_dictionary, embed_tokens)
 
         # self.length_adapter = Conv1dSubsampler(
@@ -289,7 +298,7 @@ class S2TSATEEncoder(FairseqEncoder):
         encoder_padding_mask = acoustic_encoder_out["encoder_padding_mask"][0]
 
         ctc_logit = self.acoustic_encoder.compute_ctc_logit(encoder_out)
-        ctc_prob = self.acoustic_encoder.compute_ctc_prob(encoder_out)
+        ctc_prob = self.acoustic_encoder.compute_ctc_prob(encoder_out, self.temperature)
         x = (encoder_out, ctc_prob)
 
         x, positions = self.adapter(x, encoder_padding_mask)