fix the bug of circular import

egs/mustc/mt/conf/inter.yaml

+ctc-weight: 0.2
+intermedia-ctc-weight: 0.1
+intermedia-ctc-layers: 2,4
+
+#target-ctc-weight: 0.3
+#target-ctc-layer: 6
+#target-intermedia-ctc-weight: 0.1
+#target-intermedia-ctc-layers: 2,4
+
+intermedia-adapter: league
+#intermedia-drop-prob: 0.2
+#intermedia-temperature: 5
+
+post-process: sentencepiece
\ No newline at end of file

egs/mustc/st/conf/inter.yaml

 ctc-weight: 0.2
+intermedia-ctc-weight: 0.1
 intermedia-ctc-layers: 6,9
 
+#target-ctc-weight: 0.3
+#target-ctc-layer: 6
+#target-intermedia-ctc-weight: 0.1
+#target-intermedia-ctc-layers: 2,4
+
 intermedia-adapter: league
-intermedia-ctc-weight: 0.1
 #intermedia-drop-prob: 0.2
 #intermedia-temperature: 5
 
-#target-ctc-weight: 0.5
-#target-ctc-layers: 2,4,6
-
 ctc-self-distill-weight: 0
 post-process: sentencepiece
\ No newline at end of file

fairseq/models/speech_to_text/modules/__init__.py

-from .adapter import *
-from .ctc import *
+

fairseq/models/speech_to_text/pdss2t_transformer.py

@@ -14,7 +14,7 @@ from fairseq.models import (
     register_model_architecture,
 )
 from fairseq.models.speech_to_text import S2TTransformerModel
-from fairseq.models.speech_to_text.modules import CTC, Adapter
+from fairseq.modules.speech_to_text import CTC, Adapter
 
 from fairseq.modules import (
     FairseqDropout,

fairseq/models/speech_to_text/s2t_dual.py

@@ -28,7 +28,7 @@ from fairseq.models.speech_to_text import (
     PDSS2TTransformerEncoder,
     S2TSATEEncoder,
 )
-from fairseq.models.speech_to_text.modules import Adapter, CTC
+from fairseq.modules.speech_to_text import Adapter, CTC
 from fairseq.models.transformer_s2 import (
     Embedding,
     TransformerS2Encoder,

fairseq/models/speech_to_text/s2t_sate.py

@@ -16,7 +16,7 @@ from fairseq.models.speech_to_text import (
     PDSS2TTransformerModel,
     PDSS2TTransformerEncoder,
 )
-from fairseq.models.speech_to_text.modules import Adapter, CTC
+from fairseq.modules.speech_to_text import Adapter, CTC
 from fairseq.modules import (
     FairseqDropout,
     LayerNorm,
@@ -184,7 +184,7 @@ class TextEncoder(FairseqEncoder):
         # CTC
         self.use_ctc = getattr(args, "target_ctc_weight", 0) > 0
         if self.use_ctc:
-            self.ctc_layer = args.target_ctc_layer
+            self.ctc_layer = getattr(args, "target_ctc_layer", layer_num)
             self.inter_ctc = True if self.ctc_layer != args.encoder_layers else False
             if self.inter_ctc:
                 logger.info("Target CTC loss in layer %d" % self.ctc_layer)

fairseq/models/speech_to_text/s2t_transformer.py

@@ -14,7 +14,7 @@ from fairseq.models import (
     register_model,
     register_model_architecture,
 )
-from fairseq.models.speech_to_text.modules import Adapter, CTC
+from fairseq.modules.speech_to_text import Adapter, CTC
 from fairseq.models.transformer import Embedding, TransformerDecoder
 from fairseq.modules import (
     FairseqDropout,

fairseq/models/transformer.py

@@ -18,7 +18,7 @@ from fairseq.models import (
     register_model,
     register_model_architecture,
 )
-from fairseq.models.speech_to_text.modules import Adapter, CTC
+from fairseq.modules.speech_to_text import Adapter, CTC
 from fairseq.modules import (
     AdaptiveSoftmax,
     FairseqDropout,

fairseq/modules/speech_to_text/__init__.py

 from .subsampling import *
+from .ctc import *
+from .adapter import *
\ No newline at end of file

fairseq/models/speech_to_text/modules/adapter.py → fairseq/modules/speech_to_text/adapter.py

 import logging
-import math
 
 import torch
 import torch.nn as nn