Remove input feature batchnorm..

egs/librispeech/ASR/transducer/conformer.py

@@ -56,7 +56,6 @@ class Conformer(Transformer):
         cnn_module_kernel: int = 31,
         normalize_before: bool = True,
         vgg_frontend: bool = False,
-        use_feat_batchnorm: bool = False,
     ) -> None:
         super(Conformer, self).__init__(
             num_features=num_features,
@@ -69,7 +68,6 @@ class Conformer(Transformer):
             dropout=dropout,
             normalize_before=normalize_before,
             vgg_frontend=vgg_frontend,
-            use_feat_batchnorm=use_feat_batchnorm,
         )
 
         self.encoder_pos = RelPositionalEncoding(d_model, dropout)
@@ -107,11 +105,6 @@ class Conformer(Transformer):
             - logit_lens, a tensor of shape (batch_size,) containing the number
               of frames in `logits` before padding.
         """
-        if self.use_feat_batchnorm:
-            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
-            x = self.feat_batchnorm(x)
-            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
-
         x = self.encoder_embed(x)
         x, pos_emb = self.encoder_pos(x)
         x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)

egs/librispeech/ASR/transducer/decode.py

@@ -129,7 +129,6 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
             # decoder params
             "decoder_embedding_dim": 1024,
             "num_decoder_layers": 4,
@@ -151,7 +150,6 @@ def get_encoder_model(params: AttributeDict):
         dim_feedforward=params.dim_feedforward,
         num_encoder_layers=params.num_encoder_layers,
         vgg_frontend=params.vgg_frontend,
-        use_feat_batchnorm=params.use_feat_batchnorm,
     )
     return encoder
 

egs/librispeech/ASR/transducer/export.py

@@ -119,7 +119,6 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
             # decoder params
             "decoder_embedding_dim": 1024,
             "num_decoder_layers": 4,
@@ -140,7 +139,6 @@ def get_encoder_model(params: AttributeDict):
         dim_feedforward=params.dim_feedforward,
         num_encoder_layers=params.num_encoder_layers,
         vgg_frontend=params.vgg_frontend,
-        use_feat_batchnorm=params.use_feat_batchnorm,
     )
     return encoder
 

egs/librispeech/ASR/transducer/pretrained.py

@@ -116,7 +116,6 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
             # decoder params
             "decoder_embedding_dim": 1024,
             "num_decoder_layers": 4,
@@ -137,7 +136,6 @@ def get_encoder_model(params: AttributeDict):
         dim_feedforward=params.dim_feedforward,
         num_encoder_layers=params.num_encoder_layers,
         vgg_frontend=params.vgg_frontend,
-        use_feat_batchnorm=params.use_feat_batchnorm,
     )
     return encoder
 

egs/librispeech/ASR/transducer/train.py

@@ -171,9 +171,6 @@ def get_params() -> AttributeDict:
 
         - subsampling_factor:  The subsampling factor for the model.
 
-        - use_feat_batchnorm: Whether to do batch normalization for the
-                              input features.
-
         - attention_dim: Hidden dim for multi-head attention model.
 
         - num_decoder_layers: Number of decoder layer of transformer decoder.
@@ -199,7 +196,6 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
             # decoder params
             "decoder_embedding_dim": 1024,
             "num_decoder_layers": 4,
@@ -224,7 +220,6 @@ def get_encoder_model(params: AttributeDict):
         dim_feedforward=params.dim_feedforward,
         num_encoder_layers=params.num_encoder_layers,
         vgg_frontend=params.vgg_frontend,
-        use_feat_batchnorm=params.use_feat_batchnorm,
     )
     return encoder
 

egs/librispeech/ASR/transducer/transformer.py

@@ -39,7 +39,6 @@ class Transformer(EncoderInterface):
         dropout: float = 0.1,
         normalize_before: bool = True,
         vgg_frontend: bool = False,
-        use_feat_batchnorm: bool = False,
     ) -> None:
         """
         Args:
@@ -65,13 +64,8 @@ class Transformer(EncoderInterface):
             If True, use pre-layer norm; False to use post-layer norm.
           vgg_frontend:
             True to use vgg style frontend for subsampling.
-          use_feat_batchnorm:
-            True to use batchnorm for the input layer.
         """
         super().__init__()
-        self.use_feat_batchnorm = use_feat_batchnorm
-        if use_feat_batchnorm:
-            self.feat_batchnorm = nn.BatchNorm1d(num_features)
 
         self.num_features = num_features
         self.output_dim = output_dim
@@ -131,11 +125,6 @@ class Transformer(EncoderInterface):
             - logit_lens, a tensor of shape (batch_size,) containing the number
               of frames in `logits` before padding.
         """
-        if self.use_feat_batchnorm:
-            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
-            x = self.feat_batchnorm(x)
-            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
-
         x = self.encoder_embed(x)
         x = self.encoder_pos(x)
         x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)