Remove batchnorm, weight decay, and SOS from transducer conformer encoder (#155)

* Remove batchnorm, weight decay, and SOS.

* Make --context-size configurable.

* Update results.

.github/workflows/run-pretrained-transducer-stateless.yml

@@ -74,11 +74,11 @@ jobs:
           mkdir tmp
           cd tmp
           git lfs install
-          git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22
+          git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27
           cd ..
           tree tmp
-          soxi tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/*.wav
+          soxi tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/*.wav
-          ls -lh tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/*.wav
+          ls -lh tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/*.wav
 
       - name: Run greedy search decoding
         shell: bash
@@ -87,11 +87,11 @@ jobs:
           cd egs/librispeech/ASR
           ./transducer_stateless/pretrained.py \
             --method greedy_search \
-            --checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/exp/pretrained.pt \
+            --checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/exp/pretrained.pt \
-            --bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/data/lang_bpe_500/bpe.model \
+            --bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/data/lang_bpe_500/bpe.model \
-            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/1089-134686-0001.wav \
+            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/1089-134686-0001.wav \
-            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/1221-135766-0001.wav \
+            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/1221-135766-0001.wav \
-            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/1221-135766-0002.wav
+            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/1221-135766-0002.wav
 
       - name: Run beam search decoding
         shell: bash
@@ -101,8 +101,8 @@ jobs:
           ./transducer_stateless/pretrained.py \
             --method beam_search \
             --beam-size 4 \
-            --checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/exp/pretrained.pt \
+            --checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/exp/pretrained.pt \
-            --bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/data/lang_bpe_500/bpe.model \
+            --bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/data/lang_bpe_500/bpe.model \
-            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/1089-134686-0001.wav \
+            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/1089-134686-0001.wav \
-            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/1221-135766-0001.wav \
+            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/1221-135766-0001.wav \
-            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-22/test_wavs/1221-135766-0002.wav
+            ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2021-12-27/test_wavs/1221-135766-0002.wav

README.md

@@ -84,7 +84,7 @@ The best WER using beam search with beam size 4 is:
 
 |     | test-clean | test-other |
 |-----|------------|------------|
-| WER | 2.92       | 7.37       |
+| WER | 2.83       | 7.19       |
 
 Note: No auxiliary losses are used in the training and no LMs are used
 in the decoding.

egs/librispeech/ASR/RESULTS.md

@@ -4,7 +4,7 @@
 
 #### Conformer encoder + embedding decoder
 
-Using commit `fb6a57e9e01dd8aae2af2a6b4568daad8bc8ab32`.
+Using commit `TODO`.
 
 Conformer encoder + non-current decoder. The decoder
 contains only an embedding layer and a Conv1d (with kernel size 2).
@@ -13,12 +13,8 @@ The WERs are
 
 |                           | test-clean | test-other | comment                                  |
 |---------------------------|------------|------------|------------------------------------------|
-| greedy search             | 2.99       | 7.52       | --epoch 20, --avg 10, --max-duration 100 |
+| greedy search             | 2.85       | 7.30       | --epoch 29, --avg 13, --max-duration 100 |
-| beam search (beam size 2) | 2.95       | 7.43       |                                          |
+| beam search (beam size 4) | 2.83       | 7.19       |                                          |
-| beam search (beam size 3) | 2.94       | 7.37       |                                          |
-| beam search (beam size 4) | 2.92       | 7.37       |                                          |
-| beam search (beam size 5) | 2.93       | 7.38       |                                          |
-| beam search (beam size 8) | 2.92       | 7.38       |                                          |
 
 The training command for reproducing is given below:
 
@@ -36,12 +32,12 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3"
 ```
 
 The tensorboard training log can be found at
-<https://tensorboard.dev/experiment/PsJ3LgkEQfOmzedAlYfVeg/#scalars&_smoothingWeight=0>
+<https://tensorboard.dev/experiment/Mjx7MeTgR3Oyr1yBCwjozw/>
 
 The decoding command is:
 ```
-epoch=20
+epoch=29
-avg=10
+avg=13
 
 ## greedy search
 ./transducer_stateless/decode.py \
@@ -64,7 +60,7 @@ avg=10
 
 
 #### Conformer encoder + LSTM decoder
-Using commit `TODO`.
+Using commit `8187d6236c2926500da5ee854f758e621df803cc`.
 
 Conformer encoder + LSTM decoder.
 

egs/librispeech/ASR/transducer/decode.py

@@ -396,7 +396,7 @@ def main():
     sp = spm.SentencePieceProcessor()
     sp.load(params.bpe_model)
 
-    # <blk> and <sos/eos> are defined in local/train_bpe_model.py
+    # <blk> is defined in local/train_bpe_model.py
     params.blank_id = sp.piece_to_id("<blk>")
     params.vocab_size = sp.get_piece_size()
 

egs/librispeech/ASR/transducer/export.py

@@ -194,7 +194,7 @@ def main():
     sp = spm.SentencePieceProcessor()
     sp.load(params.bpe_model)
 
-    # <blk> and <sos/eos> are defined in local/train_bpe_model.py
+    # <blk> is defined in local/train_bpe_model.py
     params.blank_id = sp.piece_to_id("<blk>")
     params.vocab_size = sp.get_piece_size()
 

egs/librispeech/ASR/transducer/pretrained.py

@@ -208,7 +208,7 @@ def main():
     sp = spm.SentencePieceProcessor()
     sp.load(params.bpe_model)
 
-    # <blk> and <sos/eos> are defined in local/train_bpe_model.py
+    # <blk> is defined in local/train_bpe_model.py
     params.blank_id = sp.piece_to_id("<blk>")
     params.vocab_size = sp.get_piece_size()
 

egs/librispeech/ASR/transducer/train.py

@@ -564,7 +564,7 @@ def run(rank, world_size, args):
     sp = spm.SentencePieceProcessor()
     sp.load(params.bpe_model)
 
-    # <blk> and <sos/eos> are defined in local/train_bpe_model.py
+    # <blk> is defined in local/train_bpe_model.py
     params.blank_id = sp.piece_to_id("<blk>")
     params.vocab_size = sp.get_piece_size()
 

egs/librispeech/ASR/transducer_stateless/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)
@@ -873,7 +866,7 @@ class ConvolutionModule(nn.Module):
             groups=channels,
             bias=bias,
         )
-        self.norm = nn.BatchNorm1d(channels)
+        self.norm = nn.LayerNorm(channels)
         self.pointwise_conv2 = nn.Conv1d(
             channels,
             channels,
@@ -903,7 +896,12 @@ class ConvolutionModule(nn.Module):
 
         # 1D Depthwise Conv
         x = self.depthwise_conv(x)
-        x = self.activation(self.norm(x))
+        # x is (batch, channels, time)
+        x = x.permute(0, 2, 1)
+        x = self.norm(x)
+        x = x.permute(0, 2, 1)
+
+        x = self.activation(x)
 
         x = self.pointwise_conv2(x)  # (batch, channel, time)
 

egs/librispeech/ASR/transducer_stateless/decode.py

@@ -70,14 +70,14 @@ def get_parser():
     parser.add_argument(
         "--epoch",
         type=int,
-        default=20,
+        default=29,
         help="It specifies the checkpoint to use for decoding."
         "Note: Epoch counts from 0.",
     )
     parser.add_argument(
         "--avg",
         type=int,
-        default=10,
+        default=13,
         help="Number of checkpoints to average. Automatically select "
         "consecutive checkpoints before the checkpoint specified by "
         "'--epoch'. ",
@@ -114,6 +114,13 @@ def get_parser():
         help="Used only when --decoding-method is beam_search",
     )
 
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
     parser.add_argument(
         "--max-sym-per-frame",
         type=int,
@@ -136,9 +143,6 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
-            # parameters for decoder
-            "context_size": 2,  # tri-gram
             "env_info": get_env_info(),
         }
     )
@@ -156,7 +160,6 @@ 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
 
@@ -393,6 +396,7 @@ def main():
     if params.decoding_method == "beam_search":
         params.suffix += f"-beam-{params.beam_size}"
     else:
+        params.suffix += f"-context-{params.context_size}"
         params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
 
     setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")

egs/librispeech/ASR/transducer_stateless/decoder.py

@@ -20,13 +20,14 @@ import torch.nn.functional as F
 
 
 class Decoder(nn.Module):
-    """This class implements the stateless decoder from the following paper:
+    """This class modifies the stateless decoder from the following paper:
 
         RNN-transducer with stateless prediction network
         https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
 
     It removes the recurrent connection from the decoder, i.e., the prediction
-    network.
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
 
     TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
     """

egs/librispeech/ASR/transducer_stateless/export.py

@@ -104,6 +104,14 @@ def get_parser():
         """,
     )
 
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
     return parser
 
 
@@ -119,9 +127,6 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
-            # parameters for decoder
-            "context_size": 2,  # tri-gram
             "env_info": get_env_info(),
         }
     )
@@ -138,7 +143,6 @@ 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_stateless/joiner.py

@@ -16,7 +16,6 @@
 
 import torch
 import torch.nn as nn
-import torch.nn.functional as F
 
 
 class Joiner(nn.Module):
@@ -48,7 +47,7 @@ class Joiner(nn.Module):
         # Now decoder_out is (N, 1, U, C)
 
         logit = encoder_out + decoder_out
-        logit = F.relu(logit)
+        logit = torch.tanh(logit)
 
         output = self.output_linear(logit)
 

egs/librispeech/ASR/transducer_stateless/pretrained.py

@@ -110,6 +110,13 @@ def get_parser():
         help="Used only when --method is beam_search",
     )
 
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
     parser.add_argument(
         "--max-sym-per-frame",
         type=int,
@@ -135,9 +142,6 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
-            # parameters for decoder
-            "context_size": 2,  # tri-gram
             "env_info": get_env_info(),
         }
     )
@@ -154,7 +158,6 @@ 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_stateless/train.py

@@ -130,6 +130,14 @@ def get_parser():
         help="The lr_factor for Noam optimizer",
     )
 
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
     return parser
 
 
@@ -171,15 +179,10 @@ 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.
 
-        - weight_decay:  The weight_decay for the optimizer.
-
         - warm_step: The warm_step for Noam optimizer.
     """
     params = AttributeDict(
@@ -201,11 +204,7 @@ def get_params() -> AttributeDict:
             "dim_feedforward": 2048,
             "num_encoder_layers": 12,
             "vgg_frontend": False,
-            "use_feat_batchnorm": True,
-            # parameters for decoder
-            "context_size": 2,  # tri-gram
             # parameters for Noam
-            "weight_decay": 1e-6,
             "warm_step": 80000,  # For the 100h subset, use 8k
             "env_info": get_env_info(),
         }
@@ -225,7 +224,6 @@ 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
 
@@ -568,7 +566,7 @@ def run(rank, world_size, args):
     sp = spm.SentencePieceProcessor()
     sp.load(params.bpe_model)
 
-    # <blk> and <sos/eos> are defined in local/train_bpe_model.py
+    # <blk> is defined in local/train_bpe_model.py
     params.blank_id = sp.piece_to_id("<blk>")
     params.vocab_size = sp.get_piece_size()
 
@@ -593,7 +591,6 @@ def run(rank, world_size, args):
         model_size=params.attention_dim,
         factor=params.lr_factor,
         warm_step=params.warm_step,
-        weight_decay=params.weight_decay,
     )
 
     if checkpoints and "optimizer" in checkpoints:

egs/librispeech/ASR/transducer_stateless/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)