from local

egs/librispeech/ASR/local/prepare_lang_bpe.py

@@ -127,7 +127,7 @@ def lexicon_to_fst_no_sil(
 
 
 def generate_lexicon(
-    model_file: str, words: List[str]
+    model_file: str, words: List[str], oov: str
 ) -> Tuple[Lexicon, Dict[str, int]]:
     """Generate a lexicon from a BPE model.
 
@@ -136,6 +136,8 @@ def generate_lexicon(
         Path to a sentencepiece model.
       words:
         A list of strings representing words.
+      oov:
+        The out of vocabulary word in lexicon.
     Returns:
       Return a tuple with two elements:
         - A dict whose keys are words and values are the corresponding
@@ -156,12 +158,9 @@ def generate_lexicon(
     for word, pieces in zip(words, words_pieces):
         lexicon.append((word, pieces))
 
-    # The OOV word is <UNK>
-    lexicon.append(("<UNK>", [sp.id_to_piece(sp.unk_id())]))
+    lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())]))
 
-    token2id: Dict[str, int] = dict()
-    for i in range(sp.vocab_size()):
-        token2id[sp.id_to_piece(i)] = i
+    token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())}
 
     return lexicon, token2id
 
@@ -176,6 +175,13 @@ def get_args():
         """,
     )
 
+    parser.add_argument(
+        "--oov",
+        type=str,
+        default="<UNK>",
+        help="The out of vocabulary word in lexicon.",
+    )
+
     parser.add_argument(
         "--debug",
         type=str2bool,
@@ -202,12 +208,13 @@ def main():
 
     words = word_sym_table.symbols
 
-    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", args.oov, "#0", "<s>", "</s>"]
+
     for w in excluded:
         if w in words:
             words.remove(w)
 
-    lexicon, token_sym_table = generate_lexicon(model_file, words)
+    lexicon, token_sym_table = generate_lexicon(model_file, words, args.oov)
 
     lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
 

egs/tedlium3/ASR/RESULTS.md

@@ -1,5 +1,88 @@
 ## Results
 
+### TedLium3 BPE training results (Conformer-CTC 2)
+
+#### [conformer_ctc2](./conformer_ctc2)
+
+See <https://github.com/k2-fsa/icefall/pull/696> for more details.
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/5NQQiqOqSqazfn4w2yeWEQ/>
+
+You can find a pretrained model and decoding results at:
+<https://huggingface.co/videodanchik/icefall-asr-tedlium3-conformer-ctc2>
+
+Number of model parameters: 101141699, i.e., 101.14 M
+
+The WERs are
+
+|                          | dev        | test        | comment             |
+|--------------------------|------------|-------------|---------------------|
+| ctc decoding             | 6.45       | 5.96        | --epoch 38 --avg 26 |
+| 1best                    | 5.92       | 5.51        | --epoch 38 --avg 26 |
+| whole lattice rescoring  | 5.96       | 5.47        | --epoch 38 --avg 26 |
+| attention decoder        | 5.60       | 5.33        | --epoch 38 --avg 26 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conformer_ctc2/train.py \
+    --world-size 4 \
+    --num-epochs 40 \
+    --exp-dir conformer_ctc2/exp \
+    --max-duration 350 \
+    --use-fp16 true
+```
+
+The decoding command is:
+```
+epoch=38
+avg=26
+
+## ctc decoding
+./conformer_ctc2/decode.py \
+  --method ctc-decoding \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+
+## 1best
+./conformer_ctc2/decode.py \
+  --method 1best \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+
+## whole lattice rescoring
+./conformer_ctc2/decode.py \
+  --method whole-lattice-rescoring \
+  --exp-dir conformer_ctc2/exp \
+  --lm-path data/lm/G_4_gram_big.pt \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+
+## attention decoder
+./conformer_ctc2/decode.py \
+  --method attention-decoder \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+```
+
 ### TedLium3 BPE training results (Pruned Transducer)
 
 #### 2022-03-21

egs/tedlium3/ASR/conformer_ctc2/asr_datamodule.py

@@ -0,0 +1,368 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2021  Xiaomi Corporation (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class TedLiumAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. TEDLium3 dev
+    and test).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset.",
+        )
+
+    def train_dataloaders(
+        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=10,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                    max_frames_mask_fraction=0.15,
+                    p=0.9,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to get Musan cuts")
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        logging.info("About to create train dataloader")
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts_test: CutSet) -> DataLoader:
+
+        logging.debug("About to create test dataset")
+        if self.args.on_the_fly_feats:
+            test = K2SpeechRecognitionDataset(
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            test = K2SpeechRecognitionDataset(
+                return_cuts=self.args.return_cuts,
+            )
+
+        test_sampler = DynamicBucketingSampler(
+            cuts_test,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=test_sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "tedlium_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_test.jsonl.gz")

egs/tedlium3/ASR/conformer_ctc2/attention.py

@@ -0,0 +1,201 @@
+# Copyright    2022  Behavox LLC.        (author: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple, Union
+
+import torch
+from scaling import ScaledLinear
+
+
+class MultiheadAttention(torch.nn.Module):
+    """Allows the model to jointly attend to information
+    from different representation subspaces. This is a modified
+    version of the original version of multihead attention
+    (see Attention Is All You Need <https://arxiv.org/abs/1706.03762>)
+    with replacement of input / output projection layers
+    with newly introduced ScaleLinear layer
+    (see https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py).
+
+    Args:
+        embed_dim:
+          total dimension of the model.
+        num_heads:
+          number of parallel attention heads. Note that embed_dim will be split
+          across num_heads, i.e. each head will have dimension (embed_dim // num_heads).
+        dropout:
+          dropout probability on attn_output_weights. (default=0.0).
+        bias:
+          if specified, adds bias to input / output projection layers (default=True).
+        add_bias_kv:
+          if specified, adds bias to the key and value sequences at dim=0 (default=False).
+        add_zero_attn:
+          if specified, adds a new batch of zeros to the key and value sequences
+          at dim=1 (default=False).
+        batch_first:
+          if True, then the input and output tensors are provided as
+          (batch, seq, feature), otherwise (seq, batch, feature) (default=False).
+
+    Examples::
+        >>> multihead_attn = MultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        bias: bool = True,
+        add_bias_kv: bool = False,
+        add_zero_attn: bool = False,
+        batch_first: bool = False,
+        device: Union[torch.device, str, None] = None,
+        dtype: Union[torch.dtype, str, None] = None,
+    ) -> None:
+
+        super().__init__()
+
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.batch_first = batch_first
+
+        if embed_dim % num_heads != 0:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads. "
+                "Got embedding dim vs number 0f heads: "
+                f"{embed_dim} vs {num_heads}"
+            )
+
+        self.head_dim = embed_dim // num_heads
+
+        self.in_proj = ScaledLinear(
+            embed_dim,
+            3 * embed_dim,
+            bias=bias,
+            device=device,
+            dtype=dtype,
+        )
+        self.out_proj = ScaledLinear(
+            embed_dim,
+            embed_dim,
+            bias=bias,
+            initial_scale=0.25,
+            device=device,
+            dtype=dtype,
+        )
+
+        if add_bias_kv:
+            self.bias_k = torch.nn.Parameter(
+                torch.empty((1, 1, embed_dim), device=device, dtype=dtype)
+            )
+            self.bias_v = torch.nn.Parameter(
+                torch.empty((1, 1, embed_dim), device=device, dtype=dtype)
+            )
+        else:
+            self.register_parameter("bias_k", None)
+            self.register_parameter("bias_v", None)
+
+        self.add_zero_attn = add_zero_attn
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        if self.bias_k is not None:
+            torch.nn.init.xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            torch.nn.init.xavier_normal_(self.bias_v)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_padding_mask: Optional[torch.Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+            query:
+              Query embeddings of shape (L, N, E_q) when batch_first=False or (N, L, E_q)
+              when batch_first=True, where L is the target sequence length, N is the batch size,
+              and E_q is the query embedding dimension embed_dim. Queries are compared against
+              key-value pairs to produce the output. See "Attention Is All You Need" for more details.
+            key:
+              Key embeddings of shape (S, N, E_k) when batch_first=False or (N, S, E_k) when
+              batch_first=True, where S is the source sequence length, N is the batch size, and
+              E_k is the key embedding dimension kdim. See "Attention Is All You Need" for more details.
+            value:
+              Value embeddings of shape (S, N, E_v) when batch_first=False or (N, S, E_v) when
+              batch_first=True, where S is the source sequence length, N is the batch size, and
+              E_v is the value embedding dimension vdim. See "Attention Is All You Need" for more details.
+            key_padding_mask:
+              If specified, a mask of shape (N, S) indicating which elements within key
+              to ignore for the purpose of attention (i.e. treat as "padding").
+              Binary and byte masks are supported. For a binary mask, a True value indicates
+              that the corresponding key value will be ignored for the purpose of attention.
+              For a byte mask, a non-zero value indicates that the corresponding key value will be ignored.
+            need_weights:
+              If specifid, returns attn_output_weights in addition to attn_outputs (default=True).
+            attn_mask:
+              If specified, a 2D or 3D mask preventing attention to certain positions. Must be of shape
+              (L, S) or (N * num_heads, L, S), where N is the batch size, L is the target sequence length,
+              and S is the source sequence length. A 2D mask will be broadcasted across the batch while
+              a 3D mask allows for a different mask for each entry in the batch.
+              Binary, byte, and float masks are supported. For a binary mask, a True value indicates
+              that the corresponding position is not allowed to attend. For a byte mask, a non-zero
+              value indicates that the corresponding position is not allowed to attend. For a float mask,
+              the mask values will be added to the attention weight.
+
+        Returns:
+            attn_output:
+              Attention outputs of shape (L, N, E) when batch_first=False or (N, L, E) when batch_first=True,
+              where L is the target sequence length, N is the batch size, and E is the embedding dimension
+              embed_dim.
+            attn_output_weights:
+              Attention output weights of shape (N, L, S), where N is the batch size, L is the target sequence
+              length, and S is the source sequence length. Only returned when need_weights=True.
+        """
+        if self.batch_first:
+            query, key, value = [x.transpose(1, 0) for x in (query, key, value)]
+
+        (
+            attn_output,
+            attn_output_weights,
+        ) = torch.nn.functional.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            self.embed_dim,
+            self.num_heads,
+            in_proj_weight=self.in_proj.get_weight(),
+            in_proj_bias=self.in_proj.get_bias(),
+            bias_k=self.bias_k,
+            bias_v=self.bias_v,
+            add_zero_attn=self.add_zero_attn,
+            dropout_p=self.dropout,
+            out_proj_weight=self.out_proj.get_weight(),
+            out_proj_bias=self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+        if self.batch_first:
+            return attn_output.transpose(1, 0), attn_output_weights
+        return attn_output, attn_output_weights

egs/tedlium3/ASR/conformer_ctc2/combiner.py

@@ -0,0 +1,244 @@
+# Copyright    2022  Behavox LLC.        (author: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List
+
+import torch
+
+
+class RandomCombine(torch.nn.Module):
+    """
+    This module combines a list of Tensors, all with the same shape, to
+    produce a single output of that same shape which, in training time,
+    is a random combination of all the inputs; but which in test time
+    will be just the last input.
+    The idea is that the list of Tensors will be a list of outputs of multiple
+    conformer layers.  This has a similar effect as iterated loss. (See:
+    DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER
+    NETWORKS).
+    """
+
+    def __init__(
+        self,
+        num_inputs: int,
+        final_weight: float = 0.5,
+        pure_prob: float = 0.5,
+        stddev: float = 2.0,
+    ) -> None:
+        """
+        Args:
+          num_inputs:
+            The number of tensor inputs, which equals the number of layers'
+            outputs that are fed into this module.  E.g. in an 18-layer neural
+            net if we output layers 16, 12, 18, num_inputs would be 3.
+          final_weight:
+            The amount of weight or probability we assign to the
+            final layer when randomly choosing layers or when choosing
+            continuous layer weights.
+          pure_prob:
+            The probability, on each frame, with which we choose
+            only a single layer to output (rather than an interpolation)
+          stddev:
+            A standard deviation that we add to log-probs for computing
+            randomized weights.
+        The method of choosing which layers, or combinations of layers, to use,
+        is conceptually as follows::
+            With probability `pure_prob`::
+               With probability `final_weight`: choose final layer,
+               Else: choose random non-final layer.
+            Else::
+               Choose initial log-weights that correspond to assigning
+               weight `final_weight` to the final layer and equal
+               weights to other layers; then add Gaussian noise
+               with variance `stddev` to these log-weights, and normalize
+               to weights (note: the average weight assigned to the
+               final layer here will not be `final_weight` if stddev>0).
+        """
+        super().__init__()
+        assert 0 <= pure_prob <= 1, pure_prob
+        assert 0 < final_weight < 1, final_weight
+        assert num_inputs >= 1, num_inputs
+
+        self.num_inputs = num_inputs
+        self.final_weight = final_weight
+        self.pure_prob = pure_prob
+        self.stddev = stddev
+
+        self.final_log_weight = (
+            torch.tensor((final_weight / (1 - final_weight)) * (self.num_inputs - 1))
+            .log()
+            .item()
+        )
+
+    def forward(self, inputs: List[torch.Tensor]) -> torch.Tensor:
+        """Forward function.
+        Args:
+          inputs:
+            A list of Tensor, e.g. from various layers of a transformer.
+            All must be the same shape, of (*, num_channels)
+        Returns:
+          A Tensor of shape (*, num_channels). In test mode
+          this is just the final input.
+        """
+        num_inputs = self.num_inputs
+        assert len(inputs) == num_inputs, f"{len(inputs)}, {num_inputs}"
+        if not self.training or torch.jit.is_scripting() or len(inputs) == 1:
+            return inputs[-1]
+
+        # Shape of weights: (*, num_inputs)
+        num_channels = inputs[0].shape[-1]
+        num_frames = inputs[0].numel() // num_channels
+
+        ndim = inputs[0].ndim
+        # stacked_inputs: (num_frames, num_channels, num_inputs)
+        stacked_inputs = torch.stack(inputs, dim=ndim).reshape(
+            (num_frames, num_channels, num_inputs)
+        )
+
+        # weights: (num_frames, num_inputs)
+        weights = self._get_random_weights(
+            inputs[0].dtype, inputs[0].device, num_frames
+        )
+
+        weights = weights.reshape(num_frames, num_inputs, 1)
+        # ans: (num_frames, num_channels, 1)
+        ans = torch.matmul(stacked_inputs, weights)
+        # ans: (*, num_channels)
+
+        ans = ans.reshape(inputs[0].shape[:-1] + (num_channels,))
+
+        return ans
+
+    def _get_random_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), such that
+          `ans.sum(dim=1)` is all ones.
+        """
+        pure_prob = self.pure_prob
+        if pure_prob == 0.0:
+            return self._get_random_mixed_weights(dtype, device, num_frames)
+        elif pure_prob == 1.0:
+            return self._get_random_pure_weights(dtype, device, num_frames)
+        else:
+            p = self._get_random_pure_weights(dtype, device, num_frames)
+            m = self._get_random_mixed_weights(dtype, device, num_frames)
+            return torch.where(
+                torch.rand(num_frames, 1, device=device) < self.pure_prob, p, m
+            )
+
+    def _get_random_pure_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A one-hot tensor of shape `(num_frames, self.num_inputs)`, with
+          exactly one weight equal to 1.0 on each frame.
+        """
+        final_prob = self.final_weight
+
+        # final contains self.num_inputs - 1 in all elements
+        final = torch.full((num_frames,), self.num_inputs - 1, device=device)
+        # nonfinal contains random integers in [0..num_inputs - 2], these are for non-final weights.
+        nonfinal = torch.randint(self.num_inputs - 1, (num_frames,), device=device)
+
+        indexes = torch.where(
+            torch.rand(num_frames, device=device) < final_prob, final, nonfinal
+        )
+        ans = torch.nn.functional.one_hot(indexes, num_classes=self.num_inputs).to(
+            dtype=dtype
+        )
+        return ans
+
+    def _get_random_mixed_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), which elements
+          in [0..1] that sum to one over the second axis, i.e.
+          `ans.sum(dim=1)` is all ones.
+        """
+        logprobs = (
+            torch.randn(num_frames, self.num_inputs, dtype=dtype, device=device)
+            * self.stddev
+        )
+        logprobs[:, -1] += self.final_log_weight
+        return logprobs.softmax(dim=1)
+
+
+def _test_random_combine(
+    final_weight: float,
+    pure_prob: float,
+    stddev: float,
+) -> None:
+    print(
+        f"_test_random_combine: final_weight={final_weight}, "
+        f"pure_prob={pure_prob}, stddev={stddev}"
+    )
+    num_inputs = 3
+    num_channels = 50
+    m = RandomCombine(
+        num_inputs=num_inputs,
+        final_weight=final_weight,
+        pure_prob=pure_prob,
+        stddev=stddev,
+    )
+
+    x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)]
+
+    y = m(x)
+    assert y.shape == x[0].shape
+    assert torch.allclose(y, x[0])  # .. since actually all ones.
+
+
+def _test_random_combine_main() -> None:
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.0)
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.3)
+    _test_random_combine(0.5, 1, 0.3)
+    _test_random_combine(0.5, 0.5, 0.3)
+
+
+if __name__ == "__main__":
+    _test_random_combine_main()

egs/tedlium3/ASR/conformer_ctc2/decode.py

@@ -0,0 +1,896 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo,
+#                                            Fangjun Kuang,
+#                                            Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from conformer import Conformer
+from train import add_model_arguments
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) ctc-greedy-search. It only use CTC output and a sentence piece
+              model for decoding. It produces the same results with ctc-decoding.
+            - (2) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (3) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (6) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-path",
+        type=str,
+        default="data/lm/G_4_gram.pt",
+        help="""The n-gram LM dir for rescoring.
+        It should contain either lm_fname.pt or lm_fname.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--result-dir",
+        type=str,
+        default="conformer_ctc2/exp/results",
+        help="Directory to store results.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+    """
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            # parameters for decoding
+            "search_beam": 15,
+            "output_beam": 8,
+            "min_active_states": 10,
+            "max_active_states": 7000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def ctc_greedy_search(
+    ctc_probs: torch.Tensor,
+    mask: torch.Tensor,
+) -> List[List[int]]:
+    """Apply CTC greedy search
+    Args:
+      ctc_probs (torch.Tensor): (batch, max_len, num_bpe)
+      mask (torch.Tensor): (batch, max_len)
+    Returns:
+      best path result
+    """
+
+    _, max_index = ctc_probs.max(2)  # (B, maxlen)
+    max_index = max_index.masked_fill_(mask, 0)  # (B, maxlen)
+
+    ret_hyps = []
+    for hyp in max_index:
+        hyp = torch.unique_consecutive(hyp)
+        hyp = hyp[hyp > 0].tolist()
+        ret_hyps.append(hyp)
+    return ret_hyps
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        unk = bpe_model.decode(bpe_model.unk_id()).strip()
+        hyps = [[w for w in s.split() if w != unk] for s in hyps]
+        key = "ctc-decoding"
+
+        return {key: hyps}
+
+    if params.method == "ctc-greedy-search":
+        hyps = ctc_greedy_search(nnet_output, memory_key_padding_mask)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(hyps)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        unk = bpe_model.decode(bpe_model.unk_id()).strip()
+        hyps = [[w for w in s.split() if w != unk] for s in hyps]
+        key = "ctc-greedy-search"
+
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<unk>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [
+            [word_table[i] for i in ids if word_table[i] != "<unk>"] for ids in hyps
+        ]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method == "nbest":
+        best_path = nbest_decoding(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            use_double_scores=params.use_double_scores,
+            nbest_scale=params.nbest_scale,
+        )
+        key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [
+            [word_table[i] for i in ids if word_table[i] != "<unk>"] for ids in hyps
+        ]
+        return {key: hyps}
+
+    assert params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "1best":
+        best_path_dict = one_best_decoding(
+            lattice=lattice,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [
+                [word_table[i] for i in ids if word_table[i] != "<unk>"] for ids in hyps
+            ]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+) -> None:
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.result_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.result_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.result_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main() -> None:
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_path = Path(args.lm_path)
+    args.result_dir = Path(args.result_dir)
+
+    args.result_dir.mkdir(exist_ok=True)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    if params.method in ("ctc-decoding", "ctc-greedy-search"):
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in ("nbest-rescoring", "whole-lattice-rescoring"):
+        assert params.lm_path.suffix in (".pt", ".txt")
+
+        if params.lm_path.is_file() and params.lm_path.suffix == ".pt":
+            logging.info(f"Loading pre-compiled {params.lm_path.name}")
+            d = torch.load(params.lm_path, map_location=device)
+            G = k2.Fsa.from_dict(d)
+        elif not params.lm_path.is_file() and params.lm_path.suffix == ".txt":
+            raise FileNotFoundError(f"No such language model file: '{params.lm_path}'")
+        else:
+            # here we pass only if LM filename ends with '.pt' and doesn't exist
+            # or if LM filename ends '.txt' and exists.
+            if (
+                not params.lm_path.is_file()
+                and params.lm_path.suffix == ".pt"
+                and not (
+                    params.lm_path.parent / f"{params.lm_path.stem}.fst.txt"
+                ).is_file()
+            ):
+                raise FileNotFoundError(
+                    f"No such language model file: '{params.lm_path}'\n"
+                    "'.fst.txt' representation of the language model was "
+                    "not found either."
+                )
+            else:
+                # whatever params.lm_path.name we got lm_name.pt or lm_name.fst.txt
+                # we are going to load lm_name.fst.txt here
+                params.lm_path = params.lm_path.parent / params.lm_path.name.replace(
+                    ".pt", ".fst.txt"
+                )
+                logging.info(f"Loading {params.lm_path.name}")
+                logging.warning("It may take 8 minutes.")
+                with open(params.lm_path) as f:
+                    first_word_disambig_id = lexicon.word_table["#0"]
+
+                    G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                    # G.aux_labels is not needed in later computations, so
+                    # remove it here.
+                    del G.aux_labels
+                    # CAUTION: The following line is crucial.
+                    # Arcs entering the back-off state have label equal to #0.
+                    # We have to change it to 0 here.
+                    G.labels[G.labels >= first_word_disambig_id] = 0
+                    # See https://github.com/k2-fsa/k2/issues/874
+                    # for why we need to set G.properties to None
+                    G.__dict__["_properties"] = None
+                    G = k2.Fsa.from_fsas([G]).to(device)
+                    G = k2.arc_sort(G)
+                    # Save a dummy value so that it can be loaded in C++.
+                    # See https://github.com/pytorch/pytorch/issues/67902
+                    # for why we need to do this.
+                    G.dummy = 1
+
+                    torch.save(
+                        G.as_dict(),
+                        params.lm_path.parent
+                        / params.lm_path.name.replace(".fst.txt", ".pt"),
+                    )
+
+        if params.method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.dim_model,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tedlium = TedLiumAsrDataModule(args)
+
+    valid_cuts = tedlium.dev_cuts()
+    test_cuts = tedlium.test_cuts()
+
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+    test_dl = tedlium.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [valid_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+# when we import add_model_arguments from train.py
+# we enforce torch.set_num_interop_threads(1) in it,
+# so we ended up with setting num_interop_threads to one
+# two times: in train.py and decode.py which cause an error,
+# that is why added an additional if statement.
+if torch.get_num_interop_threads() != 1:
+    torch.set_num_interop_threads(1)
+
+# The flag below controls whether to allow TF32 on matmul. This flag defaults to False
+# in PyTorch 1.12 and later.
+torch.backends.cuda.matmul.allow_tf32 = True
+
+if __name__ == "__main__":
+    main()

egs/tedlium3/ASR/conformer_ctc2/export.py

@@ -0,0 +1,294 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Behavox LLC (Author: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./conformer_ctc2/export.py \
+  --exp-dir ./conformer_ctc2/exp \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `conformer_ctc2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/tedlium3/ASR
+    ./conformer_ctc2/decode.py \
+        --exp-dir ./conformer_ctc2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from conformer import Conformer
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help=(
+            "Number of checkpoints to average. Automatically select "
+            "consecutive checkpoints before the checkpoint specified by "
+            "'--epoch' and '--iter'"
+        ),
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help=(
+            "Whether to load averaged model. Currently it only supports "
+            "using --epoch. If True, it would decode with the averaged model "
+            "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+            "Actually only the models with epoch number of `epoch-avg` and "
+            "`epoch` are loaded for averaging. "
+        ),
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+    """
+    # parameters for conformer
+    params = AttributeDict({"subsampling_factor": 4, "feature_dim": 80})
+    return params
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.dim_model,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                "Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/tedlium3/ASR/conformer_ctc2/label_smoothing.py

@@ -0,0 +1,109 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+class LabelSmoothingLoss(torch.nn.Module):
+    """
+    Implement the LabelSmoothingLoss proposed in the following paper
+    https://arxiv.org/pdf/1512.00567.pdf
+    (Rethinking the Inception Architecture for Computer Vision)
+
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -1,
+        label_smoothing: float = 0.1,
+        reduction: str = "sum",
+    ) -> None:
+        """
+        Args:
+          ignore_index:
+            ignored class id
+          label_smoothing:
+            smoothing rate (0.0 means the conventional cross entropy loss)
+          reduction:
+            It has the same meaning as the reduction in
+            `torch.nn.CrossEntropyLoss`. It can be one of the following three
+            values: (1) "none": No reduction will be applied. (2) "mean": the
+            mean of the output is taken. (3) "sum": the output will be summed.
+        """
+        super().__init__()
+        assert 0.0 <= label_smoothing < 1.0, f"{label_smoothing}"
+        assert reduction in ("none", "sum", "mean"), reduction
+        self.ignore_index = ignore_index
+        self.label_smoothing = label_smoothing
+        self.reduction = reduction
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Compute loss between x and target.
+
+        Args:
+          x:
+            prediction of dimension
+            (batch_size, input_length, number_of_classes).
+          target:
+            target masked with self.ignore_index of
+            dimension (batch_size, input_length).
+
+        Returns:
+          A scalar tensor containing the loss without normalization.
+        """
+        assert x.ndim == 3
+        assert target.ndim == 2
+        assert x.shape[:2] == target.shape
+        num_classes = x.size(-1)
+        x = x.reshape(-1, num_classes)
+        # Now x is of shape (N*T, C)
+
+        # We don't want to change target in-place below,
+        # so we make a copy of it here
+        target = target.clone().reshape(-1)
+
+        ignored = target == self.ignore_index
+
+        # See https://github.com/k2-fsa/icefall/issues/240
+        # and https://github.com/k2-fsa/icefall/issues/297
+        # for why we don't use target[ignored] = 0 here
+        target = torch.where(ignored, torch.zeros_like(target), target)
+
+        true_dist = torch.nn.functional.one_hot(target, num_classes=num_classes).to(x)
+
+        true_dist = (
+            true_dist * (1 - self.label_smoothing) + self.label_smoothing / num_classes
+        )
+
+        # Set the value of ignored indexes to 0
+        #
+        # See https://github.com/k2-fsa/icefall/issues/240
+        # and https://github.com/k2-fsa/icefall/issues/297
+        # for why we don't use true_dist[ignored] = 0 here
+        true_dist = torch.where(
+            ignored.unsqueeze(1).repeat(1, true_dist.shape[1]),
+            torch.zeros_like(true_dist),
+            true_dist,
+        )
+
+        loss = -1 * (torch.log_softmax(x, dim=1) * true_dist)
+        if self.reduction == "sum":
+            return loss.sum()
+        elif self.reduction == "mean":
+            return loss.sum() / (~ignored).sum()
+        else:
+            return loss.sum(dim=-1)

egs/tedlium3/ASR/conformer_ctc2/lstmp.py

@@ -0,0 +1,102 @@
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LSTMP(nn.Module):
+    """LSTM with projection.
+
+    PyTorch does not support exporting LSTM with projection to ONNX.
+    This class reimplements LSTM with projection using basic matrix-matrix
+    and matrix-vector operations. It is not intended for training.
+    """
+
+    def __init__(self, lstm: nn.LSTM):
+        """
+        Args:
+          lstm:
+            LSTM with proj_size. We support only uni-directional,
+            1-layer LSTM with projection at present.
+        """
+        super().__init__()
+        assert lstm.bidirectional is False, lstm.bidirectional
+        assert lstm.num_layers == 1, lstm.num_layers
+        assert 0 < lstm.proj_size < lstm.hidden_size, (
+            lstm.proj_size,
+            lstm.hidden_size,
+        )
+
+        assert lstm.batch_first is False, lstm.batch_first
+
+        state_dict = lstm.state_dict()
+
+        w_ih = state_dict["weight_ih_l0"]
+        w_hh = state_dict["weight_hh_l0"]
+
+        b_ih = state_dict["bias_ih_l0"]
+        b_hh = state_dict["bias_hh_l0"]
+
+        w_hr = state_dict["weight_hr_l0"]
+        self.input_size = lstm.input_size
+        self.proj_size = lstm.proj_size
+        self.hidden_size = lstm.hidden_size
+
+        self.w_ih = w_ih
+        self.w_hh = w_hh
+        self.b = b_ih + b_hh
+        self.w_hr = w_hr
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        hx: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          input:
+            A tensor of shape [T, N, hidden_size]
+          hx:
+            A tuple containing:
+              - h0: a tensor of shape (1, N, proj_size)
+              - c0: a tensor of shape (1, N, hidden_size)
+        Returns:
+          Return a tuple containing:
+            - output: a tensor of shape (T, N, proj_size).
+            - A tuple containing:
+               - h: a tensor of shape (1, N, proj_size)
+               - c: a tensor of shape (1, N, hidden_size)
+
+        """
+        x_list = input.unbind(dim=0)  # We use batch_first=False
+
+        if hx is not None:
+            h0, c0 = hx
+        else:
+            h0 = torch.zeros(1, input.size(1), self.proj_size)
+            c0 = torch.zeros(1, input.size(1), self.hidden_size)
+        h0 = h0.squeeze(0)
+        c0 = c0.squeeze(0)
+        y_list = []
+        for x in x_list:
+            gates = F.linear(x, self.w_ih, self.b) + F.linear(h0, self.w_hh)
+            i, f, g, o = gates.chunk(4, dim=1)
+
+            i = i.sigmoid()
+            f = f.sigmoid()
+            g = g.tanh()
+            o = o.sigmoid()
+
+            c = f * c0 + i * g
+            h = o * c.tanh()
+
+            h = F.linear(h, self.w_hr)
+            y_list.append(h)
+
+            c0 = c
+            h0 = h
+
+        y = torch.stack(y_list, dim=0)
+
+        return y, (h0.unsqueeze(0), c0.unsqueeze(0))

egs/tedlium3/ASR/conformer_ctc2/optim.py

@@ -0,0 +1,320 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Union
+
+import torch
+from torch.optim import Optimizer
+
+
+class Eve(Optimizer):
+    r"""
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam\: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                # Constrain the range of scalar weights
+                if p.numel() == 1:
+                    p.clamp_(min=-10, max=2)
+
+        return loss
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("initial_lr", group["lr"])
+
+        self.base_lrs = [group["initial_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            print(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+     lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                       (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25))
+
+     E.g. suggest initial-lr = 0.003 (passed to optimizer).
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+        return [x * factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = Eve(m.parameters(), lr=0.003)
+
+    scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+    print("last lr = ", scheduler.get_last_lr())
+    print("state dict = ", scheduler.state_dict())
+
+
+if __name__ == "__main__":
+    _test_eden()

egs/tedlium3/ASR/conformer_ctc2/scaling_converter.py

@@ -0,0 +1,320 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file provides functions to convert `ScaledLinear`, `ScaledConv1d`,
+`ScaledConv2d`, and `ScaledEmbedding` to their non-scaled counterparts:
+`nn.Linear`, `nn.Conv1d`, `nn.Conv2d`, and `nn.Embedding`.
+
+The scaled version are required only in the training time. It simplifies our
+life by converting them to their non-scaled version during inference.
+"""
+
+import copy
+import re
+from typing import List
+
+import torch
+import torch.nn as nn
+from lstmp import LSTMP
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledEmbedding,
+    ScaledLinear,
+    ScaledLSTM,
+)
+
+
+class NonScaledNorm(nn.Module):
+    """See BasicNorm for doc"""
+
+    def __init__(
+        self,
+        num_channels: int,
+        eps_exp: float,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+    ):
+        super().__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.eps_exp = eps_exp
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not torch.jit.is_tracing():
+            assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x * x, dim=self.channel_dim, keepdim=True) + self.eps_exp
+        ).pow(-0.5)
+        return x * scales
+
+
+def scaled_linear_to_linear(scaled_linear: ScaledLinear) -> nn.Linear:
+    """Convert an instance of ScaledLinear to nn.Linear.
+
+    Args:
+      scaled_linear:
+        The layer to be converted.
+    Returns:
+      Return a linear layer. It satisfies:
+
+        scaled_linear(x) == linear(x)
+
+      for any given input tensor `x`.
+    """
+    assert isinstance(scaled_linear, ScaledLinear), type(scaled_linear)
+
+    weight = scaled_linear.get_weight()
+    bias = scaled_linear.get_bias()
+    has_bias = bias is not None
+
+    linear = torch.nn.Linear(
+        in_features=scaled_linear.in_features,
+        out_features=scaled_linear.out_features,
+        bias=True,  # otherwise, it throws errors when converting to PNNX format
+        # device=weight.device,  # Pytorch version before v1.9.0 does not have
+        # this argument. Comment out for now, we will
+        # see if it will raise error for versions
+        # after v1.9.0
+    )
+    linear.weight.data.copy_(weight)
+
+    if has_bias:
+        linear.bias.data.copy_(bias)
+    else:
+        linear.bias.data.zero_()
+
+    return linear
+
+
+def scaled_conv1d_to_conv1d(scaled_conv1d: ScaledConv1d) -> nn.Conv1d:
+    """Convert an instance of ScaledConv1d to nn.Conv1d.
+
+    Args:
+      scaled_conv1d:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Conv1d that has the same `forward()` behavior
+      of the given `scaled_conv1d`.
+    """
+    assert isinstance(scaled_conv1d, ScaledConv1d), type(scaled_conv1d)
+
+    weight = scaled_conv1d.get_weight()
+    bias = scaled_conv1d.get_bias()
+    has_bias = bias is not None
+
+    conv1d = nn.Conv1d(
+        in_channels=scaled_conv1d.in_channels,
+        out_channels=scaled_conv1d.out_channels,
+        kernel_size=scaled_conv1d.kernel_size,
+        stride=scaled_conv1d.stride,
+        padding=scaled_conv1d.padding,
+        dilation=scaled_conv1d.dilation,
+        groups=scaled_conv1d.groups,
+        bias=scaled_conv1d.bias is not None,
+        padding_mode=scaled_conv1d.padding_mode,
+    )
+
+    conv1d.weight.data.copy_(weight)
+    if has_bias:
+        conv1d.bias.data.copy_(bias)
+
+    return conv1d
+
+
+def scaled_conv2d_to_conv2d(scaled_conv2d: ScaledConv2d) -> nn.Conv2d:
+    """Convert an instance of ScaledConv2d to nn.Conv2d.
+
+    Args:
+      scaled_conv2d:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Conv2d that has the same `forward()` behavior
+      of the given `scaled_conv2d`.
+    """
+    assert isinstance(scaled_conv2d, ScaledConv2d), type(scaled_conv2d)
+
+    weight = scaled_conv2d.get_weight()
+    bias = scaled_conv2d.get_bias()
+    has_bias = bias is not None
+
+    conv2d = nn.Conv2d(
+        in_channels=scaled_conv2d.in_channels,
+        out_channels=scaled_conv2d.out_channels,
+        kernel_size=scaled_conv2d.kernel_size,
+        stride=scaled_conv2d.stride,
+        padding=scaled_conv2d.padding,
+        dilation=scaled_conv2d.dilation,
+        groups=scaled_conv2d.groups,
+        bias=scaled_conv2d.bias is not None,
+        padding_mode=scaled_conv2d.padding_mode,
+    )
+
+    conv2d.weight.data.copy_(weight)
+    if has_bias:
+        conv2d.bias.data.copy_(bias)
+
+    return conv2d
+
+
+def scaled_embedding_to_embedding(
+    scaled_embedding: ScaledEmbedding,
+) -> nn.Embedding:
+    """Convert an instance of ScaledEmbedding to nn.Embedding.
+
+    Args:
+      scaled_embedding:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Embedding that has the same `forward()` behavior
+      of the given `scaled_embedding`.
+    """
+    assert isinstance(scaled_embedding, ScaledEmbedding), type(scaled_embedding)
+    embedding = nn.Embedding(
+        num_embeddings=scaled_embedding.num_embeddings,
+        embedding_dim=scaled_embedding.embedding_dim,
+        padding_idx=scaled_embedding.padding_idx,
+        scale_grad_by_freq=scaled_embedding.scale_grad_by_freq,
+        sparse=scaled_embedding.sparse,
+    )
+    weight = scaled_embedding.weight
+    scale = scaled_embedding.scale
+
+    embedding.weight.data.copy_(weight * scale.exp())
+
+    return embedding
+
+
+def convert_basic_norm(basic_norm: BasicNorm) -> NonScaledNorm:
+    assert isinstance(basic_norm, BasicNorm), type(BasicNorm)
+    norm = NonScaledNorm(
+        num_channels=basic_norm.num_channels,
+        eps_exp=basic_norm.eps.data.exp().item(),
+        channel_dim=basic_norm.channel_dim,
+    )
+    return norm
+
+
+def scaled_lstm_to_lstm(scaled_lstm: ScaledLSTM) -> nn.LSTM:
+    """Convert an instance of ScaledLSTM to nn.LSTM.
+
+    Args:
+      scaled_lstm:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.LSTM that has the same `forward()` behavior
+      of the given `scaled_lstm`.
+    """
+    assert isinstance(scaled_lstm, ScaledLSTM), type(scaled_lstm)
+    lstm = nn.LSTM(
+        input_size=scaled_lstm.input_size,
+        hidden_size=scaled_lstm.hidden_size,
+        num_layers=scaled_lstm.num_layers,
+        bias=scaled_lstm.bias,
+        batch_first=scaled_lstm.batch_first,
+        dropout=scaled_lstm.dropout,
+        bidirectional=scaled_lstm.bidirectional,
+        proj_size=scaled_lstm.proj_size,
+    )
+
+    assert lstm._flat_weights_names == scaled_lstm._flat_weights_names
+    for idx in range(len(scaled_lstm._flat_weights_names)):
+        scaled_weight = scaled_lstm._flat_weights[idx] * scaled_lstm._scales[idx].exp()
+        lstm._flat_weights[idx].data.copy_(scaled_weight)
+
+    return lstm
+
+
+# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule  # noqa
+# get_submodule was added to nn.Module at v1.9.0
+def get_submodule(model, target):
+    if target == "":
+        return model
+    atoms: List[str] = target.split(".")
+    mod: torch.nn.Module = model
+    for item in atoms:
+        if not hasattr(mod, item):
+            raise AttributeError(
+                mod._get_name() + " has no " "attribute `" + item + "`"
+            )
+        mod = getattr(mod, item)
+        if not isinstance(mod, torch.nn.Module):
+            raise AttributeError("`" + item + "` is not " "an nn.Module")
+    return mod
+
+
+def convert_scaled_to_non_scaled(
+    model: nn.Module,
+    inplace: bool = False,
+    is_onnx: bool = False,
+):
+    """Convert `ScaledLinear`, `ScaledConv1d`, and `ScaledConv2d`
+    in the given modle to their unscaled version `nn.Linear`, `nn.Conv1d`,
+    and `nn.Conv2d`.
+
+    Args:
+      model:
+        The model to be converted.
+      inplace:
+        If True, the input model is modified inplace.
+        If False, the input model is copied and we modify the copied version.
+      is_onnx:
+        If True, we are going to export the model to ONNX. In this case,
+        we will convert nn.LSTM with proj_size to LSTMP.
+    Return:
+      Return a model without scaled layers.
+    """
+    if not inplace:
+        model = copy.deepcopy(model)
+
+    excluded_patterns = r"(self|src)_attn\.(in|out)_proj"
+    p = re.compile(excluded_patterns)
+
+    d = {}
+    for name, m in model.named_modules():
+        if isinstance(m, ScaledLinear):
+            if p.search(name) is not None:
+                continue
+            d[name] = scaled_linear_to_linear(m)
+        elif isinstance(m, ScaledConv1d):
+            d[name] = scaled_conv1d_to_conv1d(m)
+        elif isinstance(m, ScaledConv2d):
+            d[name] = scaled_conv2d_to_conv2d(m)
+        elif isinstance(m, ScaledEmbedding):
+            d[name] = scaled_embedding_to_embedding(m)
+        elif isinstance(m, BasicNorm):
+            d[name] = convert_basic_norm(m)
+        elif isinstance(m, ScaledLSTM):
+            if is_onnx:
+                d[name] = LSTMP(scaled_lstm_to_lstm(m))
+                # See
+                # https://github.com/pytorch/pytorch/issues/47887
+                #  d[name] = torch.jit.script(LSTMP(scaled_lstm_to_lstm(m)))
+            else:
+                d[name] = scaled_lstm_to_lstm(m)
+        elif isinstance(m, ActivationBalancer):
+            d[name] = nn.Identity()
+
+    for k, v in d.items():
+        if "." in k:
+            parent, child = k.rsplit(".", maxsplit=1)
+            setattr(get_submodule(model, parent), child, v)
+        else:
+            setattr(model, k, v)
+
+    return model

egs/tedlium3/ASR/conformer_ctc2/subsampling.py

@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corporation                        (author: Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv2d,
+    ScaledLinear,
+)
+
+
+class Conv2dSubsampling(torch.nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = torch.nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x

egs/tedlium3/ASR/local/convert_transcript_words_to_bpe_ids.py

@@ -4,16 +4,18 @@
 """
 Convert a transcript based on words to a list of BPE ids.
 
-For example, if we use 2 as the encoding id of <unk>:
+For example, if we use 2 as the encoding id of <unk>
+Note: it, inserts a space token before each <unk>
 
 texts = ['this is a <unk> day']
-spm_ids = [[38, 33, 6, 2, 316]]
+spm_ids = [[38, 33, 6, 15, 2, 316]]
 
 texts = ['<unk> this is a sunny day']
-spm_ids = [[2, 38, 33, 6, 118, 11, 11, 21, 316]]
+spm_ids = [[15, 2, 38, 33, 6, 118, 11, 11, 21, 316]]
 
 texts = ['<unk>']
-spm_ids = [[2]]
+spm_ids = [[15, 2]]
+
 """
 
 import argparse
@@ -38,29 +40,27 @@ def get_args():
 
 def convert_texts_into_ids(
     texts: List[str],
-    unk_id: int,
     sp: spm.SentencePieceProcessor,
 ) -> List[List[int]]:
     """
     Args:
       texts:
         A string list of transcripts, such as ['Today is Monday', 'It's sunny'].
-      unk_id:
-        A number id for the token '<unk>'.
+      sp:
+        A sentencepiece BPE model.
     Returns:
       Return an integer list of bpe ids.
     """
     y = []
     for text in texts:
-        y_ids = []
         if "<unk>" in text:
-            text_segments = text.split("<unk>")
-            id_segments = sp.encode(text_segments, out_type=int)
+            id_segments = sp.encode(text.split("<unk>"), out_type=int)
+
+            y_ids = []
             for i in range(len(id_segments)):
-                if i != len(id_segments) - 1:
-                    y_ids.extend(id_segments[i] + [unk_id])
-                else:
-                    y_ids.extend(id_segments[i])
+                y_ids += id_segments[i]
+                if i < len(id_segments) - 1:
+                    y_ids += [sp.piece_to_id("▁"), sp.unk_id()]
         else:
             y_ids = sp.encode(text, out_type=int)
         y.append(y_ids)
@@ -70,19 +70,13 @@ def convert_texts_into_ids(
 
 def main():
     args = get_args()
-    texts = args.texts
-    bpe_model = args.bpe_model
 
     sp = spm.SentencePieceProcessor()
-    sp.load(bpe_model)
-    unk_id = sp.piece_to_id("<unk>")
+    sp.load(args.bpe_model)
 
-    y = convert_texts_into_ids(
-        texts=texts,
-        unk_id=unk_id,
-        sp=sp,
-    )
-    logging.info(f"The input texts: {texts}")
+    y = convert_texts_into_ids(texts=args.texts, sp=sp)
+
+    logging.info(f"The input texts: {args.texts}")
     logging.info(f"The encoding ids: {y}")
 
 

egs/tedlium3/ASR/local/prepare_transcripts.py

@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
-# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+# Copyright    2021  Xiaomi Corp.        (author: Mingshuang Luo)
+# Copyright    2022  Behavox LLC.        (author: Daniil Kulko)
 #
 # See ../../../../LICENSE for clarification regarding multiple authors
 #
@@ -17,68 +18,67 @@
 
 
 """
-This script takes as input supervisions json dir "data/manifests"
-consisting of supervisions_train.json and does the following:
-
-1. Generate train.text.
+This script takes input text file and removes all words
+that iclude any character out of English alphabet.
 
 """
 import argparse
 import logging
+import re
 from pathlib import Path
 
-import lhotse
-
 
 def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument(
-        "--manifests-dir",
+        "--input-text-path",
         type=str,
-        help="""Input directory.
-        """,
+        help="Input text file path.",
     )
     parser.add_argument(
-        "--lang-dir",
+        "--output-text-path",
         type=str,
-        help="""Output directory.
-        """,
+        help="Output text file path.",
     )
 
     return parser.parse_args()
 
 
-def prepare_transcripts(manifests_dir: str, lang_dir: str):
+def prepare_transcripts(input_text_path: Path, output_text_path: Path) -> None:
     """
     Args:
-      manifests_dir:
-        The manifests directory, e.g., data/manifests.
-      lang_dir:
-        The language directory, e.g., data/lang_phone.
+      input_text_path:
+        The input data text file path, e.g., data/lang/train_orig.txt.
+      output_text_path:
+        The output data text file path, e.g., data/lang/train.txt.
 
     Return:
-      The train.text in lang_dir.
+      Saved text file in output_text_path.
     """
-    texts = []
 
-    train_text = Path(lang_dir) / "train.text"
-    sups = lhotse.load_manifest(f"{manifests_dir}/tedlium_supervisions_train.jsonl.gz")
-    for s in sups:
-        texts.append(s.text)
+    foreign_chr_check = re.compile(r"[^a-z']")
 
-    with open(train_text, "w") as f:
-        for text in texts:
-            f.write(text)
-            f.write("\n")
+    logging.info(f"Loading {input_text_path.name}")
+    with open(input_text_path, "r", encoding="utf8") as f:
+        texts = {t.rstrip("\n") for t in f}
+
+    texts = {
+        " ".join([w for w in t.split() if foreign_chr_check.search(w) is None])
+        for t in texts
+    }
+
+    with open(output_text_path, "w+", encoding="utf8") as f:
+        for t in texts:
+            f.write(f"{t}\n")
 
 
-def main():
+def main() -> None:
     args = get_args()
-    manifests_dir = Path(args.manifests_dir)
-    lang_dir = Path(args.lang_dir)
+    input_text_path = Path(args.input_text_path)
+    output_text_path = Path(args.output_text_path)
 
-    logging.info("Generating train.text")
-    prepare_transcripts(manifests_dir, lang_dir)
+    logging.info(f"Generating {output_text_path.name}")
+    prepare_transcripts(input_text_path, output_text_path)
 
 
 if __name__ == "__main__":

egs/tedlium3/ASR/local/prepare_words.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+# Copyright    2022  Behavox LLC.        (authors: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input supervisions json dir "data/manifests"
+consisting of tedlium_supervisions_train.json and does the following:
+
+1. Generate words.txt.
+
+"""
+import argparse
+import logging
+import re
+from pathlib import Path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="Output directory.",
+    )
+
+    return parser.parse_args()
+
+
+def prepare_words(lang_dir: str) -> None:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang.
+
+    Return:
+      The words.txt file.
+    """
+
+    words_orig_path = Path(lang_dir) / "words_orig.txt"
+    words_path = Path(lang_dir) / "words.txt"
+
+    foreign_chr_check = re.compile(r"[^a-z']")
+
+    logging.info(f"Loading {words_orig_path.name}")
+    with open(words_orig_path, "r", encoding="utf8") as f:
+        words = {w for w_compl in f for w in w_compl.strip("-\n").split("_")}
+    words = {w for w in words if foreign_chr_check.search(w) is None and w != ""}
+    words.add("<unk>")
+    words = ["<eps>", "!SIL"] + sorted(words) + ["#0", "<s>", "</s>"]
+
+    with open(words_path, "w+", encoding="utf8") as f:
+        for idx, word in enumerate(words):
+            f.write(f"{word} {idx}\n")
+
+
+def main() -> None:
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    logging.info("Generating words.txt")
+    prepare_words(lang_dir)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/tedlium3/ASR/prepare.sh

@@ -5,7 +5,6 @@ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
 
 set -eou pipefail
 
-nj=15
 stage=0
 stop_stage=100
 
@@ -63,6 +62,13 @@ if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
     mv $dl_dir/TEDLIUM_release-3 $dl_dir/tedlium3
   fi
 
+  # Download big and small 4 gram lanuage models
+  if [ ! -d $dl_dir/lm ]; then
+    wget --continue http://kaldi-asr.org/models/5/4gram_small.arpa.gz -P $dl_dir/lm
+    wget --continue http://kaldi-asr.org/models/5/4gram_big.arpa.gz -P $dl_dir/lm
+    gzip -d $dl_dir/lm/4gram_small.arpa.gz $dl_dir/lm/4gram_big.arpa.gz
+  fi
+
   # If you have pre-downloaded it to /path/to/musan,
   # you can create a symlink
   #
@@ -100,7 +106,14 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
 
   if [ ! -e data/fbank/.tedlium3.done ]; then
     mkdir -p data/fbank
+
     python3 ./local/compute_fbank_tedlium.py
+
+    gunzip -c data/fbank/tedlium_cuts_train.jsonl.gz | shuf | \
+    gzip -c > data/fbank/tedlium_cuts_train-shuf.jsonl.gz
+    mv data/fbank/tedlium_cuts_train-shuf.jsonl.gz \
+       data/fbank/tedlium_cuts_train.jsonl.gz
+
     touch data/fbank/.tedlium3.done
   fi
 fi
@@ -115,28 +128,24 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
 fi
 
 if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
-  log "Stage 5: Prepare phone based lang"
-  lang_dir=data/lang_phone
+  log "Stage 5: Prepare BPE train data and set of words"
+  lang_dir=data/lang
   mkdir -p $lang_dir
 
-  if [ ! -f $lang_dir/train.text ]; then
+  if [ ! -f $lang_dir/train.txt ]; then
+    gunzip -c $dl_dir/tedlium3/LM/*.en.gz | sed 's: <\/s>::g' > $lang_dir/train_orig.txt
+
     ./local/prepare_transcripts.py \
-      --lang-dir $lang_dir \
-      --manifests-dir data/manifests
+      --input-text-path $lang_dir/train_orig.txt \
+      --output-text-path $lang_dir/train.txt
   fi
 
-  if [ ! -f $lang_dir/lexicon_words.txt ]; then
-    ./local/prepare_lexicon.py \
-      --lang-dir $lang_dir \
-      --manifests-dir data/manifests
-  fi
+  if [ ! -f $lang_dir/words.txt ]; then
 
-  (echo '!SIL SIL'; echo '<UNK> <UNK>'; ) |
-    cat - $lang_dir/lexicon_words.txt |
-    sort | uniq > $lang_dir/lexicon.txt
+    awk '{print $1}' $dl_dir/tedlium3/TEDLIUM.152k.dic |
+    sed 's:([0-9])::g' | sort | uniq > $lang_dir/words_orig.txt
 
-  if [ ! -f $lang_dir/L_disambig.pt ]; then
-    ./local/prepare_lang.py --lang-dir $lang_dir
+    ./local/prepare_words.py --lang-dir $lang_dir
   fi
 fi
 
@@ -148,25 +157,56 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
     mkdir -p $lang_dir
     # We reuse words.txt from phone based lexicon
     # so that the two can share G.pt later.
-    cp data/lang_phone/words.txt $lang_dir
-
-    if [ ! -f $lang_dir/transcript_words.txt ]; then
-      log "Generate data for BPE training"
-      cat data/lang_phone/train.text |
-      cut -d " " -f 2- > $lang_dir/transcript_words.txt
-      # remove the <unk> for transcript_words.txt
-      sed -i 's/ <unk>//g' $lang_dir/transcript_words.txt
-      sed -i 's/<unk> //g' $lang_dir/transcript_words.txt
-      sed -i 's/<unk>//g' $lang_dir/transcript_words.txt
-    fi
+    cp data/lang/words.txt $lang_dir
 
     ./local/train_bpe_model.py \
       --lang-dir $lang_dir \
       --vocab-size $vocab_size \
-      --transcript $lang_dir/transcript_words.txt
+      --transcript data/lang/train.txt
 
     if [ ! -f $lang_dir/L_disambig.pt ]; then
-      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir --oov "<unk>"
+    fi
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare G"
+  # We assume you have install kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_4_gram_small.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      --max-arpa-warnings=-1 \
+      $dl_dir/lm/4gram_small.arpa > data/lm/G_4_gram_small.fst.txt
+  fi
+
+  if [ ! -f data/lm/G_4_gram_big.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      --max-arpa-warnings=-1 \
+      $dl_dir/lm/4gram_big.arpa > data/lm/G_4_gram_big.fst.txt
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Compile HLG"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/HLG.pt ]; then
+      ./local/compile_hlg.py \
+        --lang-dir $lang_dir \
+        --lm G_4_gram_small
     fi
   done
 fi